Upload 58 files

.gitattributes

@@ -4,3 +4,4 @@ RobustVideoMatting/checkpoint/rvm_mobilenetv3.pth filter=lfs diff=lfs merge=lfs
 RobustVideoMatting/documentation/image/showreel.gif filter=lfs diff=lfs merge=lfs -text
 RobustVideoMatting/documentation/image/teaser.gif filter=lfs diff=lfs merge=lfs -text
 RobustVideoMatting/rvm_mobilenetv3.pth filter=lfs diff=lfs merge=lfs -text
+pifuhd/checkpoints/pifuhd.pt filter=lfs diff=lfs merge=lfs -text

pifuhd/CODE_OF_CONDUCT.md

@@ -0,0 +1,76 @@
+# Code of Conduct
+
+## Our Pledge
+
+In the interest of fostering an open and welcoming environment, we as
+contributors and maintainers pledge to make participation in our project and
+our community a harassment-free experience for everyone, regardless of age, body
+size, disability, ethnicity, sex characteristics, gender identity and expression,
+level of experience, education, socio-economic status, nationality, personal
+appearance, race, religion, or sexual identity and orientation.
+
+## Our Standards
+
+Examples of behavior that contributes to creating a positive environment
+include:
+
+* Using welcoming and inclusive language
+* Being respectful of differing viewpoints and experiences
+* Gracefully accepting constructive criticism
+* Focusing on what is best for the community
+* Showing empathy towards other community members
+
+Examples of unacceptable behavior by participants include:
+
+* The use of sexualized language or imagery and unwelcome sexual attention or
+advances
+* Trolling, insulting/derogatory comments, and personal or political attacks
+* Public or private harassment
+* Publishing others' private information, such as a physical or electronic
+address, without explicit permission
+* Other conduct which could reasonably be considered inappropriate in a
+professional setting
+
+## Our Responsibilities
+
+Project maintainers are responsible for clarifying the standards of acceptable
+behavior and are expected to take appropriate and fair corrective action in
+response to any instances of unacceptable behavior.
+
+Project maintainers have the right and responsibility to remove, edit, or
+reject comments, commits, code, wiki edits, issues, and other contributions
+that are not aligned to this Code of Conduct, or to ban temporarily or
+permanently any contributor for other behaviors that they deem inappropriate,
+threatening, offensive, or harmful.
+
+## Scope
+
+This Code of Conduct applies within all project spaces, and it also applies when
+an individual is representing the project or its community in public spaces.
+Examples of representing a project or community include using an official
+project e-mail address, posting via an official social media account, or acting
+as an appointed representative at an online or offline event. Representation of
+a project may be further defined and clarified by project maintainers.
+
+## Enforcement
+
+Instances of abusive, harassing, or otherwise unacceptable behavior may be
+reported by contacting the project team at <[email protected]>. All
+complaints will be reviewed and investigated and will result in a response that
+is deemed necessary and appropriate to the circumstances. The project team is
+obligated to maintain confidentiality with regard to the reporter of an incident.
+Further details of specific enforcement policies may be posted separately.
+
+Project maintainers who do not follow or enforce the Code of Conduct in good
+faith may face temporary or permanent repercussions as determined by other
+members of the project's leadership.
+
+## Attribution
+
+This Code of Conduct is adapted from the [Contributor Covenant][homepage], version 1.4,
+available at https://www.contributor-covenant.org/version/1/4/code-of-conduct.html
+
+[homepage]: https://www.contributor-covenant.org
+
+For answers to common questions about this code of conduct, see
+https://www.contributor-covenant.org/faq

pifuhd/CONTRIBUTING.md

@@ -0,0 +1,31 @@
+# Contributing to pifuhd
+We want to make contributing to this project as easy and transparent as
+possible.
+
+## Pull Requests
+We actively welcome your pull requests.
+
+1. Fork the repo and create your branch from `master`.
+2. If you've added code that should be tested, add tests.
+3. If you've changed APIs, update the documentation.
+4. Ensure the test suite passes.
+5. Make sure your code lints.
+6. If you haven't already, complete the Contributor License Agreement ("CLA").
+
+## Contributor License Agreement ("CLA")
+In order to accept your pull request, we need you to submit a CLA. You only need
+to do this once to work on any of Facebook's open source projects.
+
+Complete your CLA here: <https://code.facebook.com/cla>
+
+## Issues
+We use GitHub issues to track public bugs. Please ensure your description is
+clear and has sufficient instructions to be able to reproduce the issue.
+
+Facebook has a [bounty program](https://www.facebook.com/whitehat/) for the safe
+disclosure of security bugs. In those cases, please go through the process
+outlined on that page and do not file a public issue.
+
+## License
+By contributing to pifuhd, you agree that your contributions will be licensed
+under the LICENSE file in the root directory of this source tree.

pifuhd/LICENSE

@@ -0,0 +1,429 @@
+Copyright (c) Facebook, Inc. and its affiliates.
+All rights reserved.
+
+This source code is licensed under the license found in the
+LICENSE file in the root directory of this source tree.
+
+Attribution-NonCommercial 4.0 International
+
+=======================================================================
+
+Creative Commons Corporation ("Creative Commons") is not a law firm and
+does not provide legal services or legal advice. Distribution of
+Creative Commons public licenses does not create a lawyer-client or
+other relationship. Creative Commons makes its licenses and related
+information available on an "as-is" basis. Creative Commons gives no
+warranties regarding its licenses, any material licensed under their
+terms and conditions, or any related information. Creative Commons
+disclaims all liability for damages resulting from their use to the
+fullest extent possible.
+
+Using Creative Commons Public Licenses
+
+Creative Commons public licenses provide a standard set of terms and
+conditions that creators and other rights holders may use to share
+original works of authorship and other material subject to copyright
+and certain other rights specified in the public license below. The
+following considerations are for informational purposes only, are not
+exhaustive, and do not form part of our licenses.
+
+     Considerations for licensors: Our public licenses are
+     intended for use by those authorized to give the public
+     permission to use material in ways otherwise restricted by
+     copyright and certain other rights. Our licenses are
+     irrevocable. Licensors should read and understand the terms
+     and conditions of the license they choose before applying it.
+     Licensors should also secure all rights necessary before
+     applying our licenses so that the public can reuse the
+     material as expected. Licensors should clearly mark any
+     material not subject to the license. This includes other CC-
+     licensed material, or material used under an exception or
+     limitation to copyright. More considerations for licensors:
+	wiki.creativecommons.org/Considerations_for_licensors
+
+     Considerations for the public: By using one of our public
+     licenses, a licensor grants the public permission to use the
+     licensed material under specified terms and conditions. If
+     the licensor's permission is not necessary for any reason--for
+     example, because of any applicable exception or limitation to
+     copyright--then that use is not regulated by the license. Our
+     licenses grant only permissions under copyright and certain
+     other rights that a licensor has authority to grant. Use of
+     the licensed material may still be restricted for other
+     reasons, including because others have copyright or other
+     rights in the material. A licensor may make special requests,
+     such as asking that all changes be marked or described.
+     Although not required by our licenses, you are encouraged to
+     respect those requests where reasonable. More_considerations
+     for the public: 
+	wiki.creativecommons.org/Considerations_for_licensees
+
+=======================================================================
+
+Creative Commons Attribution-NonCommercial 4.0 International Public
+License
+
+By exercising the Licensed Rights (defined below), You accept and agree
+to be bound by the terms and conditions of this Creative Commons
+Attribution-NonCommercial 4.0 International Public License ("Public
+License"). To the extent this Public License may be interpreted as a
+contract, You are granted the Licensed Rights in consideration of Your
+acceptance of these terms and conditions, and the Licensor grants You
+such rights in consideration of benefits the Licensor receives from
+making the Licensed Material available under these terms and
+conditions.
+
+Section 1 -- Definitions.
+
+  a. Adapted Material means material subject to Copyright and Similar
+     Rights that is derived from or based upon the Licensed Material
+     and in which the Licensed Material is translated, altered,
+     arranged, transformed, or otherwise modified in a manner requiring
+     permission under the Copyright and Similar Rights held by the
+     Licensor. For purposes of this Public License, where the Licensed
+     Material is a musical work, performance, or sound recording,
+     Adapted Material is always produced where the Licensed Material is
+     synched in timed relation with a moving image.
+
+  b. Adapter's License means the license You apply to Your Copyright
+     and Similar Rights in Your contributions to Adapted Material in
+     accordance with the terms and conditions of this Public License.
+
+  c. Copyright and Similar Rights means copyright and/or similar rights
+     closely related to copyright including, without limitation,
+     performance, broadcast, sound recording, and Sui Generis Database
+     Rights, without regard to how the rights are labeled or
+     categorized. For purposes of this Public License, the rights
+     specified in Section 2(b)(1)-(2) are not Copyright and Similar
+     Rights.
+  d. Effective Technological Measures means those measures that, in the
+     absence of proper authority, may not be circumvented under laws
+     fulfilling obligations under Article 11 of the WIPO Copyright
+     Treaty adopted on December 20, 1996, and/or similar international
+     agreements.
+
+  e. Exceptions and Limitations means fair use, fair dealing, and/or
+     any other exception or limitation to Copyright and Similar Rights
+     that applies to Your use of the Licensed Material.
+
+  f. Licensed Material means the artistic or literary work, database,
+     or other material to which the Licensor applied this Public
+     License.
+
+  g. Licensed Rights means the rights granted to You subject to the
+     terms and conditions of this Public License, which are limited to
+     all Copyright and Similar Rights that apply to Your use of the
+     Licensed Material and that the Licensor has authority to license.
+
+  h. Licensor means the individual(s) or entity(ies) granting rights
+     under this Public License.
+
+  i. NonCommercial means not primarily intended for or directed towards
+     commercial advantage or monetary compensation. For purposes of
+     this Public License, the exchange of the Licensed Material for
+     other material subject to Copyright and Similar Rights by digital
+     file-sharing or similar means is NonCommercial provided there is
+     no payment of monetary compensation in connection with the
+     exchange.
+
+  j. Share means to provide material to the public by any means or
+     process that requires permission under the Licensed Rights, such
+     as reproduction, public display, public performance, distribution,
+     dissemination, communication, or importation, and to make material
+     available to the public including in ways that members of the
+     public may access the material from a place and at a time
+     individually chosen by them.
+
+  k. Sui Generis Database Rights means rights other than copyright
+     resulting from Directive 96/9/EC of the European Parliament and of
+     the Council of 11 March 1996 on the legal protection of databases,
+     as amended and/or succeeded, as well as other essentially
+     equivalent rights anywhere in the world.
+
+  l. You means the individual or entity exercising the Licensed Rights
+     under this Public License. Your has a corresponding meaning.
+
+Section 2 -- Scope.
+
+  a. License grant.
+
+       1. Subject to the terms and conditions of this Public License,
+          the Licensor hereby grants You a worldwide, royalty-free,
+          non-sublicensable, non-exclusive, irrevocable license to
+          exercise the Licensed Rights in the Licensed Material to:
+
+            a. reproduce and Share the Licensed Material, in whole or
+               in part, for NonCommercial purposes only; and
+
+            b. produce, reproduce, and Share Adapted Material for
+               NonCommercial purposes only.
+
+       2. Exceptions and Limitations. For the avoidance of doubt, where
+          Exceptions and Limitations apply to Your use, this Public
+          License does not apply, and You do not need to comply with
+          its terms and conditions.
+
+       3. Term. The term of this Public License is specified in Section
+          6(a).
+
+       4. Media and formats; technical modifications allowed. The
+          Licensor authorizes You to exercise the Licensed Rights in
+          all media and formats whether now known or hereafter created,
+          and to make technical modifications necessary to do so. The
+          Licensor waives and/or agrees not to assert any right or
+          authority to forbid You from making technical modifications
+          necessary to exercise the Licensed Rights, including
+          technical modifications necessary to circumvent Effective
+          Technological Measures. For purposes of this Public License,
+          simply making modifications authorized by this Section 2(a)
+          (4) never produces Adapted Material.
+
+       5. Downstream recipients.
+
+            a. Offer from the Licensor -- Licensed Material. Every
+               recipient of the Licensed Material automatically
+               receives an offer from the Licensor to exercise the
+               Licensed Rights under the terms and conditions of this
+               Public License.
+
+            b. No downstream restrictions. You may not offer or impose
+               any additional or different terms or conditions on, or
+               apply any Effective Technological Measures to, the
+               Licensed Material if doing so restricts exercise of the
+               Licensed Rights by any recipient of the Licensed
+               Material.
+
+       6. No endorsement. Nothing in this Public License constitutes or
+          may be construed as permission to assert or imply that You
+          are, or that Your use of the Licensed Material is, connected
+          with, or sponsored, endorsed, or granted official status by,
+          the Licensor or others designated to receive attribution as
+          provided in Section 3(a)(1)(A)(i).
+
+  b. Other rights.
+
+       1. Moral rights, such as the right of integrity, are not
+          licensed under this Public License, nor are publicity,
+          privacy, and/or other similar personality rights; however, to
+          the extent possible, the Licensor waives and/or agrees not to
+          assert any such rights held by the Licensor to the limited
+          extent necessary to allow You to exercise the Licensed
+          Rights, but not otherwise.
+
+       2. Patent and trademark rights are not licensed under this
+          Public License.
+
+       3. To the extent possible, the Licensor waives any right to
+          collect royalties from You for the exercise of the Licensed
+          Rights, whether directly or through a collecting society
+          under any voluntary or waivable statutory or compulsory
+          licensing scheme. In all other cases the Licensor expressly
+          reserves any right to collect such royalties, including when
+          the Licensed Material is used other than for NonCommercial
+          purposes.
+
+Section 3 -- License Conditions.
+
+Your exercise of the Licensed Rights is expressly made subject to the
+following conditions.
+
+  a. Attribution.
+
+       1. If You Share the Licensed Material (including in modified
+          form), You must:
+
+            a. retain the following if it is supplied by the Licensor
+               with the Licensed Material:
+
+                 i. identification of the creator(s) of the Licensed
+                    Material and any others designated to receive
+                    attribution, in any reasonable manner requested by
+                    the Licensor (including by pseudonym if
+                    designated);
+
+                ii. a copyright notice;
+
+               iii. a notice that refers to this Public License;
+
+                iv. a notice that refers to the disclaimer of
+                    warranties;
+
+                 v. a URI or hyperlink to the Licensed Material to the
+                    extent reasonably practicable;
+
+            b. indicate if You modified the Licensed Material and
+               retain an indication of any previous modifications; and
+
+            c. indicate the Licensed Material is licensed under this
+               Public License, and include the text of, or the URI or
+               hyperlink to, this Public License.
+
+       2. You may satisfy the conditions in Section 3(a)(1) in any
+          reasonable manner based on the medium, means, and context in
+          which You Share the Licensed Material. For example, it may be
+          reasonable to satisfy the conditions by providing a URI or
+          hyperlink to a resource that includes the required
+          information.
+
+       3. If requested by the Licensor, You must remove any of the
+          information required by Section 3(a)(1)(A) to the extent
+          reasonably practicable.
+
+       4. If You Share Adapted Material You produce, the Adapter's
+          License You apply must not prevent recipients of the Adapted
+          Material from complying with this Public License.
+
+Section 4 -- Sui Generis Database Rights.
+
+Where the Licensed Rights include Sui Generis Database Rights that
+apply to Your use of the Licensed Material:
+
+  a. for the avoidance of doubt, Section 2(a)(1) grants You the right
+     to extract, reuse, reproduce, and Share all or a substantial
+     portion of the contents of the database for NonCommercial purposes
+     only;
+
+  b. if You include all or a substantial portion of the database
+     contents in a database in which You have Sui Generis Database
+     Rights, then the database in which You have Sui Generis Database
+     Rights (but not its individual contents) is Adapted Material; and
+
+  c. You must comply with the conditions in Section 3(a) if You Share
+     all or a substantial portion of the contents of the database.
+
+For the avoidance of doubt, this Section 4 supplements and does not
+replace Your obligations under this Public License where the Licensed
+Rights include other Copyright and Similar Rights.
+
+Section 5 -- Disclaimer of Warranties and Limitation of Liability.
+
+  a. UNLESS OTHERWISE SEPARATELY UNDERTAKEN BY THE LICENSOR, TO THE
+     EXTENT POSSIBLE, THE LICENSOR OFFERS THE LICENSED MATERIAL AS-IS
+     AND AS-AVAILABLE, AND MAKES NO REPRESENTATIONS OR WARRANTIES OF
+     ANY KIND CONCERNING THE LICENSED MATERIAL, WHETHER EXPRESS,
+     IMPLIED, STATUTORY, OR OTHER. THIS INCLUDES, WITHOUT LIMITATION,
+     WARRANTIES OF TITLE, MERCHANTABILITY, FITNESS FOR A PARTICULAR
+     PURPOSE, NON-INFRINGEMENT, ABSENCE OF LATENT OR OTHER DEFECTS,
+     ACCURACY, OR THE PRESENCE OR ABSENCE OF ERRORS, WHETHER OR NOT
+     KNOWN OR DISCOVERABLE. WHERE DISCLAIMERS OF WARRANTIES ARE NOT
+     ALLOWED IN FULL OR IN PART, THIS DISCLAIMER MAY NOT APPLY TO YOU.
+
+  b. TO THE EXTENT POSSIBLE, IN NO EVENT WILL THE LICENSOR BE LIABLE
+     TO YOU ON ANY LEGAL THEORY (INCLUDING, WITHOUT LIMITATION,
+     NEGLIGENCE) OR OTHERWISE FOR ANY DIRECT, SPECIAL, INDIRECT,
+     INCIDENTAL, CONSEQUENTIAL, PUNITIVE, EXEMPLARY, OR OTHER LOSSES,
+     COSTS, EXPENSES, OR DAMAGES ARISING OUT OF THIS PUBLIC LICENSE OR
+     USE OF THE LICENSED MATERIAL, EVEN IF THE LICENSOR HAS BEEN
+     ADVISED OF THE POSSIBILITY OF SUCH LOSSES, COSTS, EXPENSES, OR
+     DAMAGES. WHERE A LIMITATION OF LIABILITY IS NOT ALLOWED IN FULL OR
+     IN PART, THIS LIMITATION MAY NOT APPLY TO YOU.
+
+  c. The disclaimer of warranties and limitation of liability provided
+     above shall be interpreted in a manner that, to the extent
+     possible, most closely approximates an absolute disclaimer and
+     waiver of all liability.
+
+Section 6 -- Term and Termination.
+
+  a. This Public License applies for the term of the Copyright and
+     Similar Rights licensed here. However, if You fail to comply with
+     this Public License, then Your rights under this Public License
+     terminate automatically.
+
+  b. Where Your right to use the Licensed Material has terminated under
+     Section 6(a), it reinstates:
+
+       1. automatically as of the date the violation is cured, provided
+          it is cured within 30 days of Your discovery of the
+          violation; or
+
+       2. upon express reinstatement by the Licensor.
+
+     For the avoidance of doubt, this Section 6(b) does not affect any
+     right the Licensor may have to seek remedies for Your violations
+     of this Public License.
+
+  c. For the avoidance of doubt, the Licensor may also offer the
+     Licensed Material under separate terms or conditions or stop
+     distributing the Licensed Material at any time; however, doing so
+     will not terminate this Public License.
+
+  d. Sections 1, 5, 6, 7, and 8 survive termination of this Public
+     License.
+
+Section 7 -- Other Terms and Conditions.
+
+  a. The Licensor shall not be bound by any additional or different
+     terms or conditions communicated by You unless expressly agreed.
+
+  b. Any arrangements, understandings, or agreements regarding the
+     Licensed Material not stated herein are separate from and
+     independent of the terms and conditions of this Public License.
+
+Section 8 -- Interpretation.
+
+  a. For the avoidance of doubt, this Public License does not, and
+     shall not be interpreted to, reduce, limit, restrict, or impose
+     conditions on any use of the Licensed Material that could lawfully
+     be made without permission under this Public License.
+
+  b. To the extent possible, if any provision of this Public License is
+     deemed unenforceable, it shall be automatically reformed to the
+     minimum extent necessary to make it enforceable. If the provision
+     cannot be reformed, it shall be severed from this Public License
+     without affecting the enforceability of the remaining terms and
+     conditions.
+
+  c. No term or condition of this Public License will be waived and no
+     failure to comply consented to unless expressly agreed to by the
+     Licensor.
+
+  d. Nothing in this Public License constitutes or may be interpreted
+     as a limitation upon, or waiver of, any privileges and immunities
+     that apply to the Licensor or You, including from the legal
+     processes of any jurisdiction or authority.
+
+=======================================================================
+
+Creative Commons is not a party to its public
+licenses. Notwithstanding, Creative Commons may elect to apply one of
+its public licenses to material it publishes and in those instances
+will be considered the “Licensor.” The text of the Creative Commons
+public licenses is dedicated to the public domain under the CC0 Public
+Domain Dedication. Except for the limited purpose of indicating that
+material is shared under a Creative Commons public license or as
+otherwise permitted by the Creative Commons policies published at
+creativecommons.org/policies, Creative Commons does not authorize the
+use of the trademark "Creative Commons" or any other trademark or logo
+of Creative Commons without its prior written consent including,
+without limitation, in connection with any unauthorized modifications
+to any of its public licenses or any other arrangements,
+understandings, or agreements concerning use of licensed material. For
+the avoidance of doubt, this paragraph does not form part of the
+public licenses.
+
+Creative Commons may be contacted at creativecommons.org.
+
+--------------------------- LICENSE FOR PIFu --------------------------------
+
+MIT License
+
+Copyright (c) 2019 Shunsuke Saito, Zeng Huang, and Ryota Natsume
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

pifuhd/README.md

@@ -0,0 +1,134 @@
+# [PIFuHD: Multi-Level Pixel-Aligned Implicit Function for High-Resolution 3D Human Digitization (CVPR 2020)](https://shunsukesaito.github.io/PIFuHD/)
+
+[![report](https://img.shields.io/badge/arxiv-report-red)](https://arxiv.org/pdf/2004.00452.pdf) [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/11z58bl3meSzo6kFqkahMa35G5jmh2Wgt?usp=sharing)
+
+News:
+* \[2020/06/15\] Demo with Google Colab (incl. visualization) is available! Please check out [#pifuhd on Twitter](https://twitter.com/search?q=%23pifuhd&src=recent_search_click&f=live) for many results tested by users!
+
+This repository contains a pytorch implementation of "Multi-Level Pixel-Aligned Implicit Function for High-Resolution 3D Human Digitization".
+
+![Teaser Image](https://shunsukesaito.github.io/PIFuHD/resources/images/pifuhd.gif)
+
+This codebase provides: 
+- test code
+- visualization code
+
+See our [blog post](https://ai.facebook.com/blog/facebook-research-at-cvpr-2020/) to learn more about our work at CVPR2020!
+
+## Demo on Google Colab
+In case you don't have an environment with GPUs to run PIFuHD, we offer Google Colab demo. You can also upload your own images and reconstruct 3D geometry together with visualization. Try our Colab demo using the following notebook: \
+[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/11z58bl3meSzo6kFqkahMa35G5jmh2Wgt)
+
+## Requirements
+- Python 3
+- [PyTorch](https://pytorch.org/) tested on 1.4.0, 1.5.0
+- json
+- PIL
+- skimage
+- tqdm
+- cv2
+
+For visualization
+- [trimesh](https://trimsh.org/) with pyembree
+- PyOpenGL
+- freeglut (use `sudo apt-get install freeglut3-dev` for ubuntu users)
+- ffmpeg
+
+Note: At least 8GB GPU memory is recommended to run PIFuHD model. 
+
+Run the following code to install all pip packages:
+```sh
+pip install -r requirements.txt 
+```
+
+
+
+## Download Pre-trained model
+
+Run the following script to download the pretrained model. The checkpoint is saved under `./checkpoints/`.
+```
+sh ./scripts/download_trained_model.sh
+```
+
+## A Quick Testing
+To process images under `./sample_images`, run the following code:
+```
+sh ./scripts/demo.sh
+```
+The resulting obj files and rendering will be saved in `./results`. You may use meshlab (http://www.meshlab.net/) to visualize the 3D mesh output (obj file). 
+
+
+## Testing
+
+1. run the following script to get joints for each image for testing (joints are used for image cropping only.). Make sure you correctly set the location of OpenPose binary. Alternatively [colab demo](https://colab.research.google.com/drive/11z58bl3meSzo6kFqkahMa35G5jmh2Wgt) provides more light-weight cropping rectange estimation without requiring openpose.
+```
+python apps/batch_openpose.py -d {openpose_root_path} -i {path_of_images} -o {path_of_images}
+```
+
+2. run the following script to run reconstruction code. Make sure to set `--input_path` to `path_of_images`, `--out_path` to where you want to dump out results, and `--ckpt_path` to the checkpoint. Note that unlike PIFu, PIFuHD doesn't require segmentation mask as input. But if you observe severe artifacts, you may try removing background with off-the-shelf tools such as [removebg](https://www.remove.bg/). If you have `{image_name}_rect.txt` instead of `{image_name}_keypoints.json`, add `--use_rect` flag. For reference, you can take a look at [colab demo](https://colab.research.google.com/drive/11z58bl3meSzo6kFqkahMa35G5jmh2Wgt).
+```
+python -m apps.simple_test
+```
+
+3. optionally, you can also remove artifacts by keeping only the biggest connected component from the mesh reconstruction with the following script. (Warning: the script will overwrite the original obj files.)
+```
+python apps/clean_mesh.py -f {path_of_objs}
+```
+
+## Visualization
+To render results with turn-table, run the following code. The rendered animation (.mp4) will be stored under `{path_of_objs}`.
+```
+python -m apps.render_turntable -f {path_of_objs} -ww {rendering_width} -hh {rendering_height} 
+# add -g for geometry rendering. default is normal visualization.
+```
+
+## Citation
+```
+@inproceedings{saito2020pifuhd,
+  title={PIFuHD: Multi-Level Pixel-Aligned Implicit Function for High-Resolution 3D Human Digitization},
+  author={Saito, Shunsuke and Simon, Tomas and Saragih, Jason and Joo, Hanbyul},
+  booktitle={CVPR},
+  year={2020}
+}
+```
+
+## Relevant Projects
+**[Monocular Real-Time Volumetric Performance Capture (ECCV 2020)](https://project-splinter.github.io/)**  
+*Ruilong Li\*, Yuliang Xiu\*, Shunsuke Saito, Zeng Huang, Kyle Olszewski, Hao Li*
+
+The first real-time PIFu by accelerating reconstruction and rendering!!
+
+**[PIFu: Pixel-Aligned Implicit Function for High-Resolution Clothed Human Digitization (ICCV 2019)](https://shunsukesaito.github.io/PIFu/)**  
+*Shunsuke Saito\*, Zeng Huang\*, Ryota Natsume\*, Shigeo Morishima, Angjoo Kanazawa, Hao Li*
+
+The original work of Pixel-Aligned Implicit Function for geometry and texture reconstruction, unifying sigle-view and multi-view methods.
+
+**[Learning to Infer Implicit Surfaces without 3d Supervision (NeurIPS 2019)](http://papers.nips.cc/paper/9039-learning-to-infer-implicit-surfaces-without-3d-supervision.pdf)**  
+*Shichen Liu, Shunsuke Saito, Weikai Chen, Hao Li*
+
+We answer to the question of "how can we learn implicit function if we don't have 3D ground truth?"
+
+**[SiCloPe: Silhouette-Based Clothed People (CVPR 2019, best paper finalist)](https://arxiv.org/pdf/1901.00049.pdf)**  
+*Ryota Natsume\*, Shunsuke Saito\*, Zeng Huang, Weikai Chen, Chongyang Ma, Hao Li, Shigeo Morishima*
+
+Our first attempt to reconstruct 3D clothed human body with texture from a single image!
+
+## Other Relevant Works
+**[ARCH: Animatable Reconstruction of Clothed Humans (CVPR 2020)](https://arxiv.org/pdf/2004.04572.pdf)**  
+*Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, Tony Tung*
+
+Learning PIFu in canonical space for animatable avatar generation!
+
+**[Robust 3D Self-portraits in Seconds (CVPR 2020)](http://www.liuyebin.com/portrait/portrait.html)**  
+*Zhe Li, Tao Yu, Chuanyu Pan, Zerong Zheng, Yebin Liu*
+
+They extend PIFu to RGBD + introduce "PIFusion" utilizing PIFu reconstruction for non-rigid fusion.
+
+**[Deep Volumetric Video from Very Sparse Multi-view Performance Capture (ECCV 2018)](http://openaccess.thecvf.com/content_ECCV_2018/papers/Zeng_Huang_Deep_Volumetric_Video_ECCV_2018_paper.pdf)**  
+*Zeng Huang, Tianye Li, Weikai Chen, Yajie Zhao, Jun Xing, Chloe LeGendre, Linjie Luo, Chongyang Ma, Hao Li*
+
+Implict surface learning for sparse view human performance capture!
+
+## License
+[CC-BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/legalcode). 
+See the [LICENSE](LICENSE) file. 

pifuhd/apps/batch_openpose.py

@@ -0,0 +1,22 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All rights reserved.
+
+import numpy as np
+
+import os
+import argparse
+
+parser = argparse.ArgumentParser()
+parser.add_argument('-d', '--openpose_dir', type=str, required=True)
+parser.add_argument('-i', '--input_root', type=str, required=True)
+parser.add_argument('-o', '--out_path', type=str, required=True)
+args = parser.parse_args()
+
+op_dir = args.openpose_dir
+input_path = args.input_root
+out_json_path = args.out_path
+
+os.makedirs(out_json_path, exist_ok=True)
+
+cmd = "cd {0}; ./build/examples/openpose/openpose.bin --image_dir {1} --write_json {2} --render_pose 2 --face --face_render 2 --hand --hand_render 2".format(op_dir, input_path, out_json_path)
+print(cmd)
+os.system(cmd)

pifuhd/apps/clean_mesh.py

@@ -0,0 +1,35 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All rights reserved.
+
+import os
+import argparse
+import trimesh
+
+
+def meshcleaning(file_dir):
+    files = sorted([f for f in os.listdir(file_dir) if '.obj' in f])
+    for i, file in enumerate(files):
+        obj_path = os.path.join(file_dir, file)
+
+        print(f"Processing: {obj_path}")
+
+        mesh = trimesh.load(obj_path)
+        cc = mesh.split(only_watertight=False)    
+
+        out_mesh = cc[0]
+        bbox = out_mesh.bounds
+        height = bbox[1,0] - bbox[0,0]
+        for c in cc:
+            bbox = c.bounds
+            if height < bbox[1,0] - bbox[0,0]:
+                height = bbox[1,0] - bbox[0,0]
+                out_mesh = c
+        
+        out_mesh.export(obj_path)
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument('-f', '--file_dir', type=str, required=True)
+    args = parser.parse_args()
+
+    meshcleaning(args.file_dir)

pifuhd/apps/recon.py

@@ -0,0 +1,224 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All rights reserved.
+
+import sys
+import os
+
+sys.path.insert(0, os.path.abspath(os.path.join(os.path.dirname(__file__), '..')))
+ROOT_PATH = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+
+import time
+import json 
+import numpy as np
+import cv2
+import random
+import torch
+import torch.nn as nn
+from tqdm import tqdm
+from torch.utils.data import DataLoader
+import matplotlib.pyplot as plt
+import matplotlib.cm as cm
+import matplotlib
+from numpy.linalg import inv
+
+from lib.options import BaseOptions
+from lib.mesh_util import save_obj_mesh_with_color, reconstruction
+from lib.data import EvalWPoseDataset, EvalDataset
+from lib.model import HGPIFuNetwNML, HGPIFuMRNet
+from lib.geometry import index
+
+from PIL import Image
+
+parser = BaseOptions()
+
+def gen_mesh(res, net, cuda, data, save_path, thresh=0.5, use_octree=True, components=False):
+    image_tensor_global = data['img_512'].to(device=cuda)
+    image_tensor = data['img'].to(device=cuda)
+    calib_tensor = data['calib'].to(device=cuda)
+
+    net.filter_global(image_tensor_global)
+    net.filter_local(image_tensor[:,None])
+
+    try:
+        if net.netG.netF is not None:
+            image_tensor_global = torch.cat([image_tensor_global, net.netG.nmlF], 0)
+        if net.netG.netB is not None:
+            image_tensor_global = torch.cat([image_tensor_global, net.netG.nmlB], 0)
+    except:
+        pass
+    
+    b_min = data['b_min']
+    b_max = data['b_max']
+    try:
+        save_img_path = save_path[:-4] + '.png'
+        save_img_list = []
+        for v in range(image_tensor_global.shape[0]):
+            save_img = (np.transpose(image_tensor_global[v].detach().cpu().numpy(), (1, 2, 0)) * 0.5 + 0.5)[:, :, ::-1] * 255.0
+            save_img_list.append(save_img)
+        save_img = np.concatenate(save_img_list, axis=1)
+        cv2.imwrite(save_img_path, save_img)
+
+        verts, faces, _, _ = reconstruction(
+            net, cuda, calib_tensor, res, b_min, b_max, thresh, use_octree=use_octree, num_samples=50000)
+        verts_tensor = torch.from_numpy(verts.T).unsqueeze(0).to(device=cuda).float()
+        # if 'calib_world' in data:
+        #     calib_world = data['calib_world'].numpy()[0]
+        #     verts = np.matmul(np.concatenate([verts, np.ones_like(verts[:,:1])],1), inv(calib_world).T)[:,:3]
+
+        color = np.zeros(verts.shape)
+        interval = 50000
+        for i in range(len(color) // interval + 1):
+            left = i * interval
+            if i == len(color) // interval:
+                right = -1
+            else:
+                right = (i + 1) * interval
+            net.calc_normal(verts_tensor[:, None, :, left:right], calib_tensor[:,None], calib_tensor)
+            nml = net.nmls.detach().cpu().numpy()[0] * 0.5 + 0.5
+            color[left:right] = nml.T
+
+        save_obj_mesh_with_color(save_path, verts, faces, color)
+    except Exception as e:
+        print(e)
+
+
+def gen_mesh_imgColor(res, net, cuda, data, save_path, thresh=0.5, use_octree=True, components=False):
+    image_tensor_global = data['img_512'].to(device=cuda)
+    image_tensor = data['img'].to(device=cuda)
+    calib_tensor = data['calib'].to(device=cuda)
+
+    net.filter_global(image_tensor_global)
+    net.filter_local(image_tensor[:,None])
+
+    try:
+        if net.netG.netF is not None:
+            image_tensor_global = torch.cat([image_tensor_global, net.netG.nmlF], 0)
+        if net.netG.netB is not None:
+            image_tensor_global = torch.cat([image_tensor_global, net.netG.nmlB], 0)
+    except:
+        pass
+
+    b_min = data['b_min']
+    b_max = data['b_max']
+    try:
+        save_img_path = save_path[:-4] + '.png'
+        save_img_list = []
+        for v in range(image_tensor_global.shape[0]):
+            save_img = (np.transpose(image_tensor_global[v].detach().cpu().numpy(), (1, 2, 0)) * 0.5 + 0.5)[:, :, ::-1] * 255.0
+            save_img_list.append(save_img)
+        save_img = np.concatenate(save_img_list, axis=1)
+        cv2.imwrite(save_img_path, save_img)
+
+        verts, faces, _, _ = reconstruction(
+            net, cuda, calib_tensor, res, b_min, b_max, thresh, use_octree=use_octree, num_samples=100000)
+        verts_tensor = torch.from_numpy(verts.T).unsqueeze(0).to(device=cuda).float()
+
+        # if this returns error, projection must be defined somewhere else
+        xyz_tensor = net.projection(verts_tensor, calib_tensor[:1])
+        uv = xyz_tensor[:, :2, :]
+        color = index(image_tensor[:1], uv).detach().cpu().numpy()[0].T
+        color = color * 0.5 + 0.5
+
+        if 'calib_world' in data:
+            calib_world = data['calib_world'].numpy()[0]
+            verts = np.matmul(np.concatenate([verts, np.ones_like(verts[:,:1])],1), inv(calib_world).T)[:,:3]
+
+        save_obj_mesh_with_color(save_path, verts, faces, color)
+
+    except Exception as e:
+        print(e)
+
+
+def recon(opt, use_rect=False):
+    # load checkpoints
+    state_dict_path = None
+    if opt.load_netMR_checkpoint_path is not None:
+        state_dict_path = opt.load_netMR_checkpoint_path
+    elif opt.resume_epoch < 0:
+        state_dict_path = '%s/%s_train_latest' % (opt.checkpoints_path, opt.name)
+        opt.resume_epoch = 0
+    else:
+        state_dict_path = '%s/%s_train_epoch_%d' % (opt.checkpoints_path, opt.name, opt.resume_epoch)
+    
+    start_id = opt.start_id
+    end_id = opt.end_id
+
+    cuda = torch.device('cuda:%d' % opt.gpu_id if torch.cuda.is_available() else 'cpu')
+
+    state_dict = None
+    if state_dict_path is not None and os.path.exists(state_dict_path):
+        print('Resuming from ', state_dict_path)
+        state_dict = torch.load(state_dict_path, map_location=cuda)    
+        print('Warning: opt is overwritten.')
+        dataroot = opt.dataroot
+        resolution = opt.resolution
+        results_path = opt.results_path
+        loadSize = opt.loadSize
+        
+        opt = state_dict['opt']
+        opt.dataroot = dataroot
+        opt.resolution = resolution
+        opt.results_path = results_path
+        opt.loadSize = loadSize
+    else:
+        raise Exception('failed loading state dict!', state_dict_path)
+    
+    # parser.print_options(opt)
+
+    if use_rect:
+        test_dataset = EvalDataset(opt)
+    else:
+        test_dataset = EvalWPoseDataset(opt)
+
+    print('test data size: ', len(test_dataset))
+    projection_mode = test_dataset.projection_mode
+
+    opt_netG = state_dict['opt_netG']
+    netG = HGPIFuNetwNML(opt_netG, projection_mode).to(device=cuda)
+    netMR = HGPIFuMRNet(opt, netG, projection_mode).to(device=cuda)
+
+    def set_eval():
+        netG.eval()
+
+    # load checkpoints
+    netMR.load_state_dict(state_dict['model_state_dict'])
+
+    os.makedirs(opt.checkpoints_path, exist_ok=True)
+    os.makedirs(opt.results_path, exist_ok=True)
+    os.makedirs('%s/%s/recon' % (opt.results_path, opt.name), exist_ok=True)
+
+    if start_id < 0:
+        start_id = 0
+    if end_id < 0:
+        end_id = len(test_dataset)
+
+    ## test
+    with torch.no_grad():
+        set_eval()
+
+        print('generate mesh (test) ...')
+        for i in tqdm(range(start_id, end_id)):
+            if i >= len(test_dataset):
+                break
+            
+            # for multi-person processing, set it to False
+            if True:
+                test_data = test_dataset[i]
+
+                save_path = '%s/%s/recon/result_%s_%d.obj' % (opt.results_path, opt.name, test_data['name'], opt.resolution)
+
+                print(save_path)
+                gen_mesh(opt.resolution, netMR, cuda, test_data, save_path, components=opt.use_compose)
+            else:
+                for j in range(test_dataset.get_n_person(i)):
+                    test_dataset.person_id = j
+                    test_data = test_dataset[i]
+                    save_path = '%s/%s/recon/result_%s_%d.obj' % (opt.results_path, opt.name, test_data['name'], j)
+                    gen_mesh(opt.resolution, netMR, cuda, test_data, save_path, components=opt.use_compose)
+
+def reconWrapper(args=None, use_rect=False):
+    opt = parser.parse(args)
+    recon(opt, use_rect)
+
+if __name__ == '__main__':
+    reconWrapper()
+  

pifuhd/apps/render_turntable.py

@@ -0,0 +1,129 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All rights reserved.
+
+import math
+import numpy as np
+import sys
+import os
+
+sys.path.insert(0, os.path.abspath(os.path.join(os.path.dirname(__file__), '..')))
+ROOT_PATH = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+
+from lib.render.mesh import load_obj_mesh, compute_normal
+from lib.render.camera import Camera
+from lib.render.gl.geo_render import GeoRender
+from lib.render.gl.color_render import ColorRender
+import trimesh
+
+import cv2
+import os
+import argparse
+
+width = 512
+height = 512
+
+def make_rotate(rx, ry, rz):
+    
+    sinX = np.sin(rx)
+    sinY = np.sin(ry)
+    sinZ = np.sin(rz)
+
+    cosX = np.cos(rx)
+    cosY = np.cos(ry)
+    cosZ = np.cos(rz)
+
+    Rx = np.zeros((3,3))
+    Rx[0, 0] = 1.0
+    Rx[1, 1] = cosX
+    Rx[1, 2] = -sinX
+    Rx[2, 1] = sinX
+    Rx[2, 2] = cosX
+
+    Ry = np.zeros((3,3))
+    Ry[0, 0] = cosY
+    Ry[0, 2] = sinY
+    Ry[1, 1] = 1.0
+    Ry[2, 0] = -sinY
+    Ry[2, 2] = cosY
+
+    Rz = np.zeros((3,3))
+    Rz[0, 0] = cosZ
+    Rz[0, 1] = -sinZ
+    Rz[1, 0] = sinZ
+    Rz[1, 1] = cosZ
+    Rz[2, 2] = 1.0
+
+    R = np.matmul(np.matmul(Rz,Ry),Rx)
+    return R
+
+parser = argparse.ArgumentParser()
+parser.add_argument('-f', '--file_dir', type=str, required=True)
+parser.add_argument('-ww', '--width', type=int, default=512)
+parser.add_argument('-hh', '--height', type=int, default=512)
+parser.add_argument('-g', '--geo_render', action='store_true', help='default is normal rendering')
+
+args = parser.parse_args()
+
+if args.geo_render:
+    renderer = GeoRender(width=args.width, height=args.height)
+else:
+    renderer = ColorRender(width=args.width, height=args.height)
+cam = Camera(width=1.0, height=args.height/args.width)
+cam.ortho_ratio = 1.2
+cam.near = -100
+cam.far = 10
+
+obj_files = []
+for (root,dirs,files) in os.walk(args.file_dir, topdown=True): 
+    for file in files:
+        if '.obj' in file:
+            obj_files.append(os.path.join(root, file))
+print(obj_files)
+
+R = make_rotate(math.radians(180),0,0)
+
+for i, obj_path in enumerate(obj_files):
+
+    print(obj_path)
+    obj_file = obj_path.split('/')[-1]
+    obj_root = obj_path.replace(obj_file,'')
+    file_name = obj_file[:-4]
+
+    if not os.path.exists(obj_path):
+        continue    
+    mesh = trimesh.load(obj_path)
+    vertices = mesh.vertices
+    faces = mesh.faces
+
+    # vertices = np.matmul(vertices, R.T)
+    bbox_max = vertices.max(0)
+    bbox_min = vertices.min(0)
+
+    # notice that original scale is discarded to render with the same size
+    vertices -= 0.5 * (bbox_max + bbox_min)[None,:]
+    vertices /= bbox_max[1] - bbox_min[1]
+
+    normals = compute_normal(vertices, faces)
+    
+    if args.geo_render:
+        renderer.set_mesh(vertices, faces, normals, faces)
+    else:
+        renderer.set_mesh(vertices, faces, 0.5*normals+0.5, faces) 
+        
+    cnt = 0
+    for j in range(0, 361, 2):
+        cam.center = np.array([0, 0, 0])
+        cam.eye = np.array([2.0*math.sin(math.radians(j)), 0, 2.0*math.cos(math.radians(j))]) + cam.center
+
+        renderer.set_camera(cam)
+        renderer.display()
+        
+        img = renderer.get_color(0)
+        img = cv2.cvtColor(img, cv2.COLOR_RGBA2BGRA)
+
+        cv2.imwrite(os.path.join(obj_root, 'rot_%04d.png' % cnt), 255*img)
+        cnt += 1
+
+    cmd = 'ffmpeg -framerate 30 -i ' + obj_root + '/rot_%04d.png -vcodec libx264 -y -pix_fmt yuv420p -refs 16 ' + os.path.join(obj_root, file_name + '.mp4')
+    os.system(cmd)
+    cmd = 'rm %s/rot_*.png' % obj_root
+    os.system(cmd)

pifuhd/apps/simple_test.py

@@ -0,0 +1,31 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All rights reserved.
+
+
+from .recon import reconWrapper
+import argparse
+
+
+###############################################################################################
+##                   Setting
+###############################################################################################
+parser = argparse.ArgumentParser()
+parser.add_argument('-i', '--input_path', type=str, default='./sample_images')
+parser.add_argument('-o', '--out_path', type=str, default='./results')
+parser.add_argument('-c', '--ckpt_path', type=str, default='./checkpoints/pifuhd.pt')
+parser.add_argument('-r', '--resolution', type=int, default=512)
+parser.add_argument('--use_rect', action='store_true', help='use rectangle for cropping')
+args = parser.parse_args()
+###############################################################################################
+##                   Upper PIFu
+###############################################################################################
+
+resolution = str(args.resolution)
+
+start_id = -1
+end_id = -1
+cmd = ['--dataroot', args.input_path, '--results_path', args.out_path,\
+       '--loadSize', '1024', '--resolution', resolution, '--load_netMR_checkpoint_path', \
+       args.ckpt_path,\
+       '--start_id', '%d' % start_id, '--end_id', '%d' % end_id]
+reconWrapper(cmd, args.use_rect)
+

pifuhd/checkpoints/pifuhd.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:891637c65f122c7efe212ebba7ae4581f86d75c92fdd5894096fe32f562175e9
+size 1548375177

pifuhd/data/RenderPeople_all.csv

@@ -0,0 +1,500 @@
+rp_aaron_posed_001
+rp_aaron_posed_005
+rp_aaron_posed_012
+rp_aaron_posed_019
+rp_alisha_posed_001
+rp_alison_posed_018
+rp_alison_posed_027
+rp_alison_posed_034
+rp_alison_posed_035
+rp_alvin_posed_002
+rp_alvin_posed_007
+rp_alvin_posed_011
+rp_alvin_posed_014
+rp_amber_posed_003
+rp_amber_posed_006
+rp_amber_posed_011
+rp_amber_posed_018
+rp_amber_posed_023
+rp_amber_posed_028
+rp_amir_posed_002
+rp_amir_posed_003
+rp_amir_posed_004
+rp_andrew_posed_001
+rp_andrew_posed_002
+rp_andrew_posed_004
+rp_aneko_posed_001
+rp_aneko_posed_004
+rp_aneko_posed_022
+rp_anna_posed_001
+rp_anna_posed_007
+rp_anna_posed_008
+rp_antonia_posed_003
+rp_antonia_posed_005
+rp_antonia_posed_020
+rp_ashley_posed_001
+rp_ashley_posed_003
+rp_ashley_posed_005
+rp_beatrice_posed_007
+rp_beatrice_posed_017
+rp_beatrice_posed_024
+rp_beatrice_posed_030
+rp_beatrice_posed_036
+rp_belle_posed_004
+rp_belle_posed_005
+rp_belle_posed_006
+rp_ben_posed_001
+rp_ben_posed_003
+rp_ben_posed_007
+rp_berkan_posed_001
+rp_berkan_posed_002
+rp_berkan_posed_003
+rp_bianca_posed_010
+rp_bianca_posed_019
+rp_bianca_posed_025
+rp_bob_posed_001
+rp_bob_posed_002
+rp_brad_posed_002
+rp_brad_posed_005
+rp_brad_posed_006
+rp_brandon_posed_006
+rp_brandon_posed_013
+rp_brandon_posed_019
+rp_brandon_posed_024
+rp_caren_posed_008
+rp_caren_posed_009
+rp_caren_posed_011
+rp_caren_posed_020
+rp_carina_posed_006
+rp_carina_posed_007
+rp_carla_posed_003
+rp_carla_posed_012
+rp_carla_posed_018
+rp_carla_posed_026
+rp_celina_posed_002
+rp_celina_posed_003
+rp_celina_posed_004
+rp_chen_posed_001
+rp_chen_posed_003
+rp_chen_posed_006
+rp_chloe_posed_002
+rp_chloe_posed_003
+rp_christine_posed_001
+rp_christine_posed_014
+rp_christine_posed_025
+rp_christopher_posed_001
+rp_christopher_posed_003
+rp_christopher_posed_005
+rp_christopher_posed_008
+rp_cindy_posed_004
+rp_cindy_posed_009
+rp_cindy_posed_019
+rp_claudia_posed_003
+rp_claudia_posed_017
+rp_claudia_posed_019
+rp_claudia_posed_026
+rp_corey_posed_004
+rp_corey_posed_010
+rp_corey_posed_017
+rp_corey_posed_026
+rp_cornell_posed_003
+rp_cornell_posed_009
+rp_cornell_posed_014
+rp_dana_posed_001
+rp_daniel_posed_001
+rp_daniel_posed_003
+rp_daniel_posed_004
+rp_dave_posed_004
+rp_dave_posed_005
+rp_debra_posed_002
+rp_debra_posed_013
+rp_debra_posed_017
+rp_dennis_posed_018
+rp_dennis_posed_024
+rp_dennis_posed_031
+rp_dennis_posed_035
+rp_elena_posed_006
+rp_elena_posed_010
+rp_elena_posed_013
+rp_elias_posed_001
+rp_elias_posed_005
+rp_elias_posed_012
+rp_elizabeth_posed_001
+rp_elizabeth_posed_002
+rp_elizabeth_posed_004
+rp_ellie_posed_005
+rp_ellie_posed_014
+rp_ellie_posed_016
+rp_emily_posed_008
+rp_emily_posed_009
+rp_emily_posed_010
+rp_emily_posed_016
+rp_emily_posed_023
+rp_emily_posed_027
+rp_emma_posed_006
+rp_emma_posed_010
+rp_emma_posed_012
+rp_emma_posed_020
+rp_emma_posed_025
+rp_emma_posed_032
+rp_eric_posed_013
+rp_eric_posed_027
+rp_eric_posed_029
+rp_eric_posed_039
+rp_ethan_posed_001
+rp_ethan_posed_004
+rp_ethan_posed_006
+rp_ethan_posed_016
+rp_ethan_posed_021
+rp_ethan_posed_023
+rp_eve_posed_001
+rp_fabienne_posed_003
+rp_fabienne_posed_011
+rp_fabienne_posed_014
+rp_felice_posed_002
+rp_felice_posed_004
+rp_felice_posed_021
+rp_felice_posed_024
+rp_felix_posed_001
+rp_fernanda_posed_016
+rp_fernanda_posed_026
+rp_fernanda_posed_032
+rp_fernanda_posed_035
+rp_fiona_posed_001
+rp_fiona_posed_003
+rp_fiona_posed_013
+rp_fiona_posed_017
+rp_frank_posed_002
+rp_frank_posed_015
+rp_frank_posed_020
+rp_frank_posed_023
+rp_george_posed_002
+rp_george_posed_003
+rp_grace_posed_001
+rp_grace_posed_002
+rp_grace_posed_004
+rp_hannah_posed_007
+rp_hannah_posed_009
+rp_helen_posed_010
+rp_helen_posed_012
+rp_helen_posed_027
+rp_helen_posed_034
+rp_helen_posed_035
+rp_henry_posed_001
+rp_henry_posed_002
+rp_henry_posed_010
+rp_henry_posed_015
+rp_holly_posed_002
+rp_holly_posed_004
+rp_holly_posed_006
+rp_holly_posed_010
+rp_holly_posed_014
+rp_jacob_posed_001
+rp_jacob_posed_003
+rp_janett_posed_001
+rp_janett_posed_007
+rp_janett_posed_019
+rp_janett_posed_029
+rp_janna_posed_019
+rp_janna_posed_026
+rp_janna_posed_046
+rp_janna_posed_048
+rp_jason_posed_002
+rp_jason_posed_008
+rp_jason_posed_011
+rp_jasper_posed_007
+rp_jasper_posed_010
+rp_jennifer_posed_003
+rp_jennifer_posed_004
+rp_jeremy_posed_001
+rp_jessica_posed_020
+rp_jessica_posed_034
+rp_jessica_posed_040
+rp_jessica_posed_053
+rp_joel_posed_003
+rp_joel_posed_009
+rp_joel_posed_011
+rp_joel_posed_016
+rp_johnny_posed_001
+rp_johnny_posed_002
+rp_john_posed_001
+rp_john_posed_002
+rp_joko_posed_001
+rp_joko_posed_009
+rp_joko_posed_013
+rp_joko_posed_016
+rp_joscha_posed_003
+rp_joscha_posed_004
+rp_joscha_posed_006
+rp_joyce_posed_010
+rp_joyce_posed_015
+rp_joyce_posed_029
+rp_judy_posed_001
+rp_judy_posed_002
+rp_julia_posed_037
+rp_julia_posed_041
+rp_julia_posed_043
+rp_julia_posed_050
+rp_kai_posed_009
+rp_kai_posed_014
+rp_kai_posed_023
+rp_kai_posed_024
+rp_kati_posed_003
+rp_kati_posed_006
+rp_kati_posed_012
+rp_kelly_posed_003
+rp_kent_posed_002
+rp_kent_posed_007
+rp_kent_posed_009
+rp_kim_posed_001
+rp_kim_posed_003
+rp_kim_posed_004
+rp_koji_posed_001
+rp_koji_posed_003
+rp_koji_posed_010
+rp_koji_posed_014
+rp_kumar_posed_001
+rp_kumar_posed_005
+rp_kumar_posed_009
+rp_kumar_posed_020
+rp_kylie_posed_010
+rp_kylie_posed_012
+rp_kylie_posed_014
+rp_kylie_posed_016
+rp_kylie_posed_017
+rp_kylie_posed_018
+rp_lars_posed_001
+rp_laura_posed_002
+rp_lee_posed_001
+rp_lee_posed_002
+rp_lee_posed_005
+rp_lee_posed_006
+rp_lena_posed_001
+rp_leo_posed_002
+rp_liam_posed_001
+rp_lina_posed_002
+rp_lina_posed_003
+rp_lina_posed_006
+rp_liz_posed_002
+rp_liz_posed_003
+rp_liz_posed_005
+rp_liz_posed_006
+rp_luisa_posed_001
+rp_luisa_posed_011
+rp_luisa_posed_020
+rp_luisa_posed_023
+rp_luisa_posed_024
+rp_luke_posed_002
+rp_luke_posed_006
+rp_luke_posed_009
+rp_maria_posed_005
+rp_maria_posed_007
+rp_maria_posed_008
+rp_marie_posed_001
+rp_mark_posed_002
+rp_mark_posed_003
+rp_mark_posed_004
+rp_mark_posed_005
+rp_marleen_posed_001
+rp_marleen_posed_002
+rp_marleen_posed_003
+rp_martha_posed_002
+rp_maurice_posed_001
+rp_maurice_posed_004
+rp_maurice_posed_006
+rp_maurice_posed_010
+rp_maurice_posed_012
+rp_maurice_posed_015
+rp_max_posed_005
+rp_max_posed_009
+rp_maya_posed_006
+rp_maya_posed_014
+rp_maya_posed_019
+rp_maya_posed_027
+rp_mei_posed_002
+rp_mei_posed_009
+rp_mei_posed_011
+rp_melanie_posed_004
+rp_melanie_posed_011
+rp_melanie_posed_012
+rp_melanie_posed_023
+rp_melinda_posed_003
+rp_melinda_posed_006
+rp_melinda_posed_008
+rp_mia_posed_001
+rp_michael_posed_006
+rp_michael_posed_009
+rp_michael_posed_018
+rp_michael_posed_025
+rp_mika_posed_001
+rp_mira_posed_006
+rp_mira_posed_016
+rp_mira_posed_019
+rp_mira_posed_024
+rp_nagy_posed_001
+rp_naomi_posed_001
+rp_naomi_posed_005
+rp_naomi_posed_009
+rp_naomi_posed_030
+rp_nathan_posed_002
+rp_nathan_posed_003
+rp_nathan_posed_011
+rp_nathan_posed_018
+rp_nathan_posed_024
+rp_nathan_posed_026
+rp_nina_posed_002
+rp_nina_posed_004
+rp_nina_posed_009
+rp_noah_posed_005
+rp_noah_posed_011
+rp_noah_posed_012
+rp_oliver_posed_003
+rp_oliver_posed_012
+rp_oliver_posed_023
+rp_oliver_posed_027
+rp_olivia_posed_002
+rp_olivia_posed_014
+rp_olivia_posed_022
+rp_olivia_posed_025
+rp_paige_posed_005
+rp_paige_posed_009
+rp_paige_posed_011
+rp_pamela_posed_003
+rp_pamela_posed_011
+rp_pamela_posed_012
+rp_patrick_posed_001
+rp_percy_posed_003
+rp_percy_posed_010
+rp_percy_posed_011
+rp_peter_posed_001
+rp_peter_posed_002
+rp_petra_posed_006
+rp_petra_posed_007
+rp_petra_posed_009
+rp_petra_posed_020
+rp_philip_posed_001
+rp_philip_posed_009
+rp_philip_posed_028
+rp_philip_posed_030
+rp_philip_posed_032
+rp_pierre_posed_002
+rp_pierre_posed_005
+rp_pierre_posed_006
+rp_ralph_posed_009
+rp_ralph_posed_010
+rp_ralph_posed_013
+rp_ralph_posed_015
+rp_ramon_posed_003
+rp_ramon_posed_005
+rp_ramon_posed_007
+rp_ray_posed_001
+rp_ray_posed_003
+rp_ray_posed_004
+rp_ray_posed_011
+rp_ricarda_posed_004
+rp_ricarda_posed_009
+rp_ricarda_posed_014
+rp_ricarda_posed_029
+rp_richard_posed_005
+rp_richard_posed_008
+rp_richard_posed_031
+rp_richard_posed_037
+rp_rick_posed_002
+rp_rick_posed_009
+rp_rick_posed_011
+rp_rick_posed_014
+rp_roberta_posed_005
+rp_roberta_posed_011
+rp_roberta_posed_023
+rp_roberta_posed_032
+rp_rosy_posed_009
+rp_rosy_posed_013
+rp_rosy_posed_016
+rp_ryo_posed_005
+rp_ryo_posed_009
+rp_ryo_posed_015
+rp_ryo_posed_016
+rp_saki_posed_012
+rp_saki_posed_020
+rp_saki_posed_024
+rp_saki_posed_025
+rp_saki_posed_030
+rp_samantha_posed_004
+rp_samantha_posed_006
+rp_samantha_posed_007
+rp_samantha_posed_009
+rp_samantha_posed_011
+rp_sarah_posed_001
+rp_sarah_posed_003
+rp_sarah_posed_004
+rp_scott_posed_006
+rp_scott_posed_013
+rp_scott_posed_029
+rp_scott_posed_036
+rp_seiko_posed_002
+rp_seiko_posed_016
+rp_seiko_posed_019
+rp_serena_posed_002
+rp_serena_posed_005
+rp_serena_posed_011
+rp_serena_posed_022
+rp_shawn_posed_002
+rp_shawn_posed_011
+rp_shawn_posed_026
+rp_shawn_posed_032
+rp_shawn_posed_037
+rp_sheila_posed_011
+rp_sheila_posed_012
+rp_sheila_posed_019
+rp_sheila_posed_032
+rp_sophia_posed_012
+rp_sophia_posed_022
+rp_sophia_posed_032
+rp_sophia_posed_039
+rp_stephanie_posed_003
+rp_stephanie_posed_009
+rp_stephanie_posed_012
+rp_stephen_posed_015
+rp_stephen_posed_031
+rp_stephen_posed_034
+rp_stephen_posed_045
+rp_steve_posed_001
+rp_sydney_posed_004
+rp_sydney_posed_014
+rp_sydney_posed_024
+rp_sydney_posed_030
+rp_tanja_posed_003
+rp_tanja_posed_007
+rp_tanja_posed_012
+rp_tanja_posed_017
+rp_thomas_posed_005
+rp_thomas_posed_009
+rp_thomas_posed_017
+rp_thomas_posed_021
+rp_tilda_posed_001
+rp_tilda_posed_002
+rp_tilda_posed_003
+rp_tim_posed_001
+rp_tony_posed_002
+rp_toshiro_posed_006
+rp_toshiro_posed_012
+rp_toshiro_posed_015
+rp_toshiro_posed_036
+rp_toshiro_posed_038
+rp_tyler_posed_003
+rp_tyler_posed_008
+rp_tyler_posed_009
+rp_tyler_posed_013
+rp_tyrone_posed_002
+rp_tyrone_posed_003
+rp_tyrone_posed_005
+rp_vanessa_posed_001
+rp_victoria_posed_003
+rp_victoria_posed_004
+rp_victoria_posed_008
+rp_wendy_posed_001
+rp_wendy_posed_002
+rp_yasmin_posed_002
+rp_yasmin_posed_015
+rp_yasmin_posed_030
+rp_zoey_posed_001

pifuhd/data/RenderPeople_test.csv

@@ -0,0 +1,50 @@
+rp_ben_posed_001
+rp_ben_posed_003
+rp_ben_posed_007
+rp_caren_posed_008
+rp_caren_posed_009
+rp_caren_posed_011
+rp_caren_posed_020
+rp_christopher_posed_001
+rp_christopher_posed_003
+rp_christopher_posed_005
+rp_christopher_posed_008
+rp_emily_posed_008
+rp_emily_posed_009
+rp_emily_posed_010
+rp_emily_posed_016
+rp_emily_posed_023
+rp_emily_posed_027
+rp_eve_posed_001
+rp_george_posed_002
+rp_george_posed_003
+rp_laura_posed_002
+rp_mira_posed_006
+rp_mira_posed_016
+rp_mira_posed_019
+rp_mira_posed_024
+rp_oliver_posed_003
+rp_oliver_posed_012
+rp_oliver_posed_023
+rp_oliver_posed_027
+rp_pamela_posed_003
+rp_pamela_posed_011
+rp_pamela_posed_012
+rp_saki_posed_012
+rp_saki_posed_020
+rp_saki_posed_024
+rp_saki_posed_025
+rp_saki_posed_030
+rp_serena_posed_002
+rp_serena_posed_005
+rp_serena_posed_011
+rp_serena_posed_022
+rp_sheila_posed_011
+rp_sheila_posed_012
+rp_sheila_posed_019
+rp_sheila_posed_032
+rp_stephen_posed_015
+rp_stephen_posed_031
+rp_stephen_posed_034
+rp_stephen_posed_045
+rp_zoey_posed_001

pifuhd/generate_normals.py

@@ -0,0 +1,200 @@
+import torch
+from torch.utils.data import Dataset
+import torchvision.transforms as transforms
+from lib.networks import define_G
+from glob import glob
+import argparse
+import os
+import os.path as osp
+import cv2, pdb
+from tqdm import tqdm
+import numpy as np
+from PIL import Image
+parser = argparse.ArgumentParser(description='neu video body rec')
+parser.add_argument('--gid',default=0,type=int,metavar='ID',
+					help='gpu id')
+parser.add_argument('--imgpath',default=None,metavar='M',
+					help='config file')
+args = parser.parse_args()
+
+
+def crop_image(img, rect):
+	x, y, w, h = rect
+
+	left = abs(x) if x < 0 else 0
+	top = abs(y) if y < 0 else 0
+	right = abs(img.shape[1]-(x+w)) if x + w >= img.shape[1] else 0
+	bottom = abs(img.shape[0]-(y+h)) if y + h >= img.shape[0] else 0
+	
+	if img.shape[2] == 4:
+		color = [0, 0, 0, 0]
+	else:
+		color = [0, 0, 0]
+	new_img = cv2.copyMakeBorder(img, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color)
+
+	x = x + left
+	y = y + top
+	# pdb.set_trace()
+	return new_img[y:(y+h),x:(x+w),:]
+
+
+
+class EvalDataset(Dataset):
+	def __init__(self, root):
+		self.root=root
+		# self.img_files=[osp.join(self.root,f) for f in os.listdir(self.root) 
+		#   				if f.split('.')[-1] in ['png', 'jpeg', 'jpg', 'PNG', 'JPG', 'JPEG'] 
+		# 				and osp.exists(osp.join(self.root,f.replace('.%s' % (f.split('.')[-1]), '_rect.txt')))]
+		self.img_files=[osp.join(self.root,f) for f in os.listdir(self.root) 
+		  				if f.split('.')[-1] in ['png', 'jpeg', 'jpg', 'PNG', 'JPG', 'JPEG']] 
+						
+		
+		self.img_files.sort(key=lambda x: int(osp.basename(x).split('.')[0]))
+		# self.img_files=sorted([osp.join(self.root,f) for f in ['0.png'] if f.split('.')[-1] in ['png', 'jpeg', 'jpg', 'PNG', 'JPG', 'JPEG'] and osp.exists(osp.join(self.root,f.replace('.%s' % (f.split('.')[-1]), '_rect.txt')))])
+
+		self.to_tensor = transforms.Compose([
+			transforms.ToTensor(),
+			transforms.Normalize((0.5,0.5,0.5),(0.5,0.5,0.5))
+		])
+		self.person_id=0
+	def __len__(self):
+		return len(self.img_files)
+	
+	def get_item(self, index):
+		# index = 386
+		img_path = self.img_files[index]
+		# rect_path = self.img_files[index].replace('.%s' % (self.img_files[index].split('.')[-1]), '_rect.txt')
+		mask_path = self.img_files[index].replace('/imgs/','/masks/')[:-3]+'png'
+
+		# Name
+		img_name = os.path.splitext(os.path.basename(img_path))[0]
+		# pdb.set_trace()
+		im = cv2.imread(img_path, cv2.IMREAD_UNCHANGED)
+		# print(mask_path)
+		if osp.isfile(mask_path):
+			mask=cv2.imread(mask_path)
+			bg=~(mask>0).all(-1)
+			im[bg]=np.zeros(im.shape[-1],dtype=im.dtype)
+		else:
+			bg=None
+		H,W=im.shape[:2]
+		if im.shape[2] == 4:
+			im = im / 255.0
+			im[:,:,:3] /= im[:,:,3:] + 1e-8
+			im = im[:,:,3:] * im[:,:,:3] + 0.5 * (1.0 - im[:,:,3:])
+			im = (255.0 * im).astype(np.uint8)
+		h, w = im.shape[:2]
+
+		# rects = np.loadtxt(rect_path, dtype=np.float64)
+		# rects[-2:] *= 1.1
+		# rects = rects.astype(np.int32)
+		# pdb.set_trace()
+		# TODO: change the rects using mask, get x, y, w, h
+		# get the y1,y2,x1,x2
+		rects = self.mask_to_bbox(mask)
+		
+		# pdb.set_trace()
+
+		if len(rects.shape) == 1:
+			rects = rects[None]
+			pid=0
+		else:
+			max_len=0
+			pid=-1
+			for ind,rect in enumerate(rects):
+				cur_len=(rect[-2]+rect[-1])//2
+				if max_len<cur_len:
+					max_len=cur_len
+					pid=ind
+		# pid = min(rects.shape[0]-1, self.person_id)
+
+		rect = rects[pid].tolist()
+		im = crop_image(im, rect)
+		im_512 = cv2.resize(im, (512, 512))
+		image_512 = Image.fromarray(im_512[:,:,::-1]).convert('RGB')
+
+		# image
+		image_512 = self.to_tensor(image_512)
+		return (img_name,image_512.unsqueeze(0),bg,H,W,rect)
+
+	def __getitem__(self, index):
+		return self.get_item(index)
+
+	def mask_to_bbox(self, mask):
+		y_ind, x_ind = np.where((mask > 0).all(-1))
+		y1, y2, x1, x2 = y_ind.min(), y_ind.max(), x_ind.min(), x_ind.max()
+		h, w = y2 - y1, x2 - x1
+		h_, w_ = 1.05 * h, 1.05 * w
+		y_, x_ = y1 - (h_ - h) / 2, x1 - (w_ - w) / 2
+		length = max(h_, w_)
+		rects = np.array([x_, y_, length, length], dtype=np.int32)
+		return rects
+
+
+
+device=torch.device(args.gid)
+
+
+# save_root=osp.normpath(osp.join(args.imgpath,osp.pardir,'normals'))
+# os.makedirs(save_root,exist_ok=True)
+
+netF=define_G(3, 3, 64, "global", 4, 9, 1, 3, "instance")
+
+weights={}
+for k,v in torch.load('checkpoints/pifuhd.pt',map_location='cpu')['model_state_dict'].items():
+	if k[:10]=='netG.netF.':
+		weights[k[10:]]=v
+
+netF.load_state_dict(weights)
+
+netF=netF.to(device)
+
+netF.eval()
+cids=[temp for temp in os.listdir(args.imgpath) if osp.isdir(osp.join(args.imgpath,temp)) and temp.isdigit()]
+
+if len(cids)==0:
+	cids=['.']
+for fold in cids:
+	save_root=osp.normpath(osp.join(args.imgpath,osp.pardir,'normals',fold))
+	print(save_root)
+	os.makedirs(save_root,exist_ok=True)
+	dataset=EvalDataset(osp.normpath(osp.join(args.imgpath,fold)))
+	writer=None
+	with torch.no_grad():	
+		for i in tqdm(range(len(dataset))):
+			# pdb.set_trace()
+			img_name,img,bg,H,W,rect=dataset[i]
+			if writer is None:
+				writer=cv2.VideoWriter(osp.join(save_root,'video.avi'),cv2.VideoWriter.fourcc('M','J','P','G'),30.,(W,H))		
+			x,y,w,h=[float(tmp) for tmp in rect]
+			# cv2.imwrite('test.png',((np.transpose(img.numpy()[0],(1,2,0))*0.5+0.5)[:,:,::-1]*255.0).astype(np.uint8))
+			
+			img=img.to(device)
+			nml=netF.forward(img) # normal: already normalized between [-1,1]
+
+			gridH,gridW=torch.meshgrid([torch.arange(H).float().to(device),torch.arange(W).float().to(device)])
+			coords=torch.stack([gridW,gridH]).permute(1,2,0).unsqueeze(0)
+			#pdb.set_trace()
+			# Here is do what? grid_sample says the coords should be in [-1, 1], but here is in [-2, 2]
+			coords[...,0] = 2.0 * (coords[...,0] - x)/w - 1.0
+			coords[...,1] = 2.0 * (coords[...,1] - y)/h - 1.0
+			# * coording to normalized coordinates, billinearly compute the normals 
+			nml=torch.nn.functional.grid_sample(nml,coords,mode='bilinear', padding_mode='zeros', align_corners=True)
+
+			unvalid_mask=(torch.norm(nml,dim=1)<0.0001).detach().cpu().numpy()[0]
+			nml=nml.detach().cpu().numpy()[0]
+			# save normal map as rgb images
+			nml=(np.transpose(nml,(1,2,0))*0.5+0.5)[:,:,::-1]*255.0 # *0.5 -> [-0.5,0.5] + 0.5 -> [0,1] * 255 -> [0,255]
+			if unvalid_mask.sum()>0:
+				nml[unvalid_mask]=0.
+			# print(osp.join(save_root,img_name,'.png'))
+			if bg is not None:
+				nml[bg]=0.
+			# if (unvalid_mask*(~bg)).sum()>0:
+			# 	print(i)
+			cv2.imwrite(osp.join(save_root,img_name+'.png'),nml.astype(np.uint8))
+			writer.write(nml.astype(np.uint8))
+
+	if writer is not None:
+		writer.release()
+print('done.')

pifuhd/lib/__init__.py

@@ -0,0 +1 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All rights reserved.

pifuhd/lib/colab_util.py

@@ -0,0 +1,140 @@
+'''
+MIT License
+
+Copyright (c) 2019 Shunsuke Saito, Zeng Huang, and Ryota Natsume
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+'''
+import io
+import os
+import torch
+from skimage.io import imread
+import numpy as np
+import cv2
+from tqdm import tqdm_notebook as tqdm
+import base64
+from IPython.display import HTML
+
+# Util function for loading meshes
+from pytorch3d.io import load_objs_as_meshes
+
+from IPython.display import HTML
+from base64 import b64encode
+
+# Data structures and functions for rendering
+from pytorch3d.structures import Meshes
+from pytorch3d.renderer import (
+    look_at_view_transform,
+    OpenGLOrthographicCameras, 
+    PointLights, 
+    DirectionalLights, 
+    Materials, 
+    RasterizationSettings, 
+    MeshRenderer, 
+    MeshRasterizer,  
+    HardPhongShader,
+    TexturesVertex
+)
+
+def set_renderer():
+    # Setup
+    device = torch.device("cuda:0")
+    torch.cuda.set_device(device)
+
+    # Initialize an OpenGL perspective camera.
+    R, T = look_at_view_transform(2.0, 0, 180) 
+    cameras = OpenGLOrthographicCameras(device=device, R=R, T=T)
+
+    raster_settings = RasterizationSettings(
+        image_size=512, 
+        blur_radius=0.0, 
+        faces_per_pixel=1, 
+        bin_size = None, 
+        max_faces_per_bin = None
+    )
+
+    lights = PointLights(device=device, location=((2.0, 2.0, 2.0),))
+
+    renderer = MeshRenderer(
+        rasterizer=MeshRasterizer(
+            cameras=cameras, 
+            raster_settings=raster_settings
+        ),
+        shader=HardPhongShader(
+            device=device, 
+            cameras=cameras,
+            lights=lights
+        )
+    )
+    return renderer
+
+def get_verts_rgb_colors(obj_path):
+  rgb_colors = []
+
+  f = open(obj_path)
+  lines = f.readlines()
+  for line in lines:
+    ls = line.split(' ')
+    if len(ls) == 7:
+      rgb_colors.append(ls[-3:])
+
+  return np.array(rgb_colors, dtype='float32')[None, :, :]
+
+def generate_video_from_obj(obj_path, image_path, video_path, renderer):
+    input_image = cv2.imread(image_path)
+    input_image = input_image[:,:input_image.shape[1]//3]
+    input_image = cv2.resize(input_image, (512,512))
+
+    # Setup
+    device = torch.device("cuda:0")
+    torch.cuda.set_device(device)
+
+    # Load obj file
+    verts_rgb_colors = get_verts_rgb_colors(obj_path)
+    verts_rgb_colors = torch.from_numpy(verts_rgb_colors).to(device)
+    textures = TexturesVertex(verts_features=verts_rgb_colors)
+    # wo_textures = TexturesVertex(verts_features=torch.ones_like(verts_rgb_colors)*0.75)
+
+    # Load obj
+    mesh = load_objs_as_meshes([obj_path], device=device)
+
+    # Set mesh
+    vers = mesh._verts_list
+    faces = mesh._faces_list
+    mesh_w_tex = Meshes(vers, faces, textures)
+    # mesh_wo_tex = Meshes(vers, faces, wo_textures)
+
+    # create VideoWriter
+    fourcc = cv2. VideoWriter_fourcc(*'MP4V')
+    out = cv2.VideoWriter(video_path, fourcc, 20.0, (1024,512))
+
+    for i in tqdm(range(90)):
+        R, T = look_at_view_transform(1.8, 0, i*4, device=device)
+        images_w_tex = renderer(mesh_w_tex, R=R, T=T)
+        images_w_tex = np.clip(images_w_tex[0, ..., :3].cpu().numpy(), 0.0, 1.0)[:, :, ::-1] * 255
+        # images_wo_tex = renderer(mesh_wo_tex, R=R, T=T)
+        # images_wo_tex = np.clip(images_wo_tex[0, ..., :3].cpu().numpy(), 0.0, 1.0)[:, :, ::-1] * 255
+        image = np.concatenate([input_image, images_w_tex], axis=1)
+        out.write(image.astype('uint8'))
+    out.release()
+
+def video(path):
+    mp4 = open(path,'rb').read()
+    data_url = "data:video/mp4;base64," + b64encode(mp4).decode()
+    return HTML('<video width=500 controls loop> <source src="%s" type="video/mp4"></video>' % data_url)

pifuhd/lib/data/EvalDataset.py

@@ -0,0 +1,131 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All rights reserved.
+
+import os
+import random
+
+import numpy as np 
+from PIL import Image, ImageOps
+from PIL.ImageFilter import GaussianBlur
+import cv2
+import torch
+import json
+
+from torch.utils.data import Dataset
+import torchvision.transforms as transforms
+
+def crop_image(img, rect):
+    x, y, w, h = rect
+
+    left = abs(x) if x < 0 else 0
+    top = abs(y) if y < 0 else 0
+    right = abs(img.shape[1]-(x+w)) if x + w >= img.shape[1] else 0
+    bottom = abs(img.shape[0]-(y+h)) if y + h >= img.shape[0] else 0
+    
+    if img.shape[2] == 4:
+        color = [0, 0, 0, 0]
+    else:
+        color = [0, 0, 0]
+    new_img = cv2.copyMakeBorder(img, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color)
+
+    x = x + left
+    y = y + top
+
+    return new_img[y:(y+h),x:(x+w),:]
+
+class EvalDataset(Dataset):
+    @staticmethod
+    def modify_commandline_options(parser, is_train):
+        return parser
+
+    def __init__(self, opt, projection='orthogonal'):
+        self.opt = opt
+        self.projection_mode = projection
+
+        self.root = self.opt.dataroot
+        self.img_files = sorted([os.path.join(self.root,f) for f in os.listdir(self.root) if f.split('.')[-1] in ['png', 'jpeg', 'jpg', 'PNG', 'JPG', 'JPEG'] and os.path.exists(os.path.join(self.root,f.replace('.%s' % (f.split('.')[-1]), '_rect.txt')))])
+        self.IMG = os.path.join(self.root)
+
+        self.phase = 'val'
+        self.load_size = self.opt.loadSize
+
+        # PIL to tensor
+        self.to_tensor = transforms.Compose([
+            transforms.ToTensor(),
+            transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
+        ])
+
+        # only used in case of multi-person processing
+        self.person_id = 0
+
+    def __len__(self):
+        return len(self.img_files)
+
+    def get_n_person(self, index):
+        rect_path = self.img_files[index].replace('.%s' % (self.img_files[index].split('.')[-1]), '_rect.txt')
+        rects = np.loadtxt(rect_path, dtype=np.int32)
+
+        return rects.shape[0] if len(rects.shape) == 2 else 1
+
+    def get_item(self, index):
+        img_path = self.img_files[index]
+        rect_path = self.img_files[index].replace('.%s' % (self.img_files[index].split('.')[-1]), '_rect.txt')
+        # Name
+        img_name = os.path.splitext(os.path.basename(img_path))[0]
+
+        im = cv2.imread(img_path, cv2.IMREAD_UNCHANGED)
+        if im.shape[2] == 4:
+            im = im / 255.0
+            im[:,:,:3] /= im[:,:,3:] + 1e-8
+            im = im[:,:,3:] * im[:,:,:3] + 0.5 * (1.0 - im[:,:,3:])
+            im = (255.0 * im).astype(np.uint8)
+        h, w = im.shape[:2]
+        
+        intrinsic = np.identity(4)
+
+        trans_mat = np.identity(4)
+
+        rects = np.loadtxt(rect_path, dtype=np.int32)
+        if len(rects.shape) == 1:
+            rects = rects[None]
+        pid = min(rects.shape[0]-1, self.person_id)
+
+        rect = rects[pid].tolist()
+        im = crop_image(im, rect)
+
+        scale_im2ndc = 1.0 / float(w // 2)
+        scale = w / rect[2]
+        trans_mat *= scale
+        trans_mat[3,3] = 1.0
+        trans_mat[0, 3] = -scale*(rect[0] + rect[2]//2 - w//2) * scale_im2ndc
+        trans_mat[1, 3] = scale*(rect[1] + rect[3]//2 - h//2) * scale_im2ndc
+        
+        intrinsic = np.matmul(trans_mat, intrinsic)
+        im_512 = cv2.resize(im, (512, 512))
+        im = cv2.resize(im, (self.load_size, self.load_size))
+
+        image_512 = Image.fromarray(im_512[:,:,::-1]).convert('RGB')
+        image = Image.fromarray(im[:,:,::-1]).convert('RGB')
+        
+        B_MIN = np.array([-1, -1, -1])
+        B_MAX = np.array([1, 1, 1])
+        projection_matrix = np.identity(4)
+        projection_matrix[1, 1] = -1
+        calib = torch.Tensor(projection_matrix).float()
+
+        calib_world = torch.Tensor(intrinsic).float()
+
+        # image
+        image_512 = self.to_tensor(image_512)
+        image = self.to_tensor(image)
+        return {
+            'name': img_name,
+            'img': image.unsqueeze(0),
+            'img_512': image_512.unsqueeze(0),
+            'calib': calib.unsqueeze(0),
+            'calib_world': calib_world.unsqueeze(0),
+            'b_min': B_MIN,
+            'b_max': B_MAX,
+        }
+
+    def __getitem__(self, index):
+        return self.get_item(index)

pifuhd/lib/data/EvalWPoseDataset.py

@@ -0,0 +1,282 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All rights reserved.
+
+import os
+import random
+
+import numpy as np 
+from PIL import Image, ImageOps
+from PIL.ImageFilter import GaussianBlur
+import cv2
+import torch
+import json
+
+from torch.utils.data import Dataset
+import torchvision.transforms as transforms
+
+def crop_image(img, rect):
+    x, y, w, h = rect
+
+    left = abs(x) if x < 0 else 0
+    top = abs(y) if y < 0 else 0
+    right = abs(img.shape[1]-(x+w)) if x + w >= img.shape[1] else 0
+    bottom = abs(img.shape[0]-(y+h)) if y + h >= img.shape[0] else 0
+    
+    if img.shape[2] == 4:
+        color = [0, 0, 0, 0]
+    else:
+        color = [0, 0, 0]
+    new_img = cv2.copyMakeBorder(img, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color)
+
+    x = x + left
+    y = y + top
+
+    return new_img[y:(y+h),x:(x+w),:]
+
+def face_crop(pts):
+    flag = pts[:,2] > 0.2
+
+    mshoulder = pts[1,:2]
+    rear = pts[18,:2]
+    lear = pts[17,:2]
+    nose = pts[0,:2]
+
+    center = np.copy(mshoulder)
+    center[1] = min(nose[1] if flag[0] else 1e8, lear[1] if flag[17] else 1e8, rear[1] if flag[18] else 1e8)
+
+    ps = []
+    pts_id = [0, 15, 16, 17, 18]
+    cnt = 0
+    for i in pts_id:
+        if flag[i]:
+            ps.append(pts[i,:2])
+            if i in [17, 18]:
+                cnt += 1
+
+    ps = np.stack(ps, 0)
+    if ps.shape[0] <= 1:
+        raise IOError('key points are not properly set')
+    if ps.shape[0] <= 3 and cnt != 2:
+        center = ps[-1]
+    else:
+        center = ps.mean(0)
+    radius = int(1.4*np.max(np.sqrt(((ps - center[None,:])**2).reshape(-1,2).sum(0))))
+
+
+    # radius = np.max(np.sqrt(((center[None] - np.stack([]))**2).sum(0))
+    # radius = int(1.0*abs(center[1] - mshoulder[1]))
+    center = center.astype(np.int)
+
+    x1 = center[0] - radius
+    x2 = center[0] + radius
+    y1 = center[1] - radius
+    y2 = center[1] + radius
+
+    return (x1, y1, x2-x1, y2-y1)
+
+def upperbody_crop(pts):
+    flag = pts[:,2] > 0.2
+
+    mshoulder = pts[1,:2]
+    ps = []
+    pts_id = [8]
+    for i in pts_id:
+        if flag[i]:
+            ps.append(pts[i,:2])
+
+    center = mshoulder
+    if len(ps) == 1:
+        ps = np.stack(ps, 0)
+        radius = int(0.8*np.max(np.sqrt(((ps - center[None,:])**2).reshape(-1,2).sum(1))))
+    else:
+        ps = []
+        pts_id = [0, 2, 5]
+        ratio = [0.4, 0.3, 0.3]
+        for i in pts_id:
+            if flag[i]:
+                ps.append(pts[i,:2])
+        ps = np.stack(ps, 0)
+        radius = int(0.8*np.max(np.sqrt(((ps - center[None,:])**2).reshape(-1,2).sum(1)) / np.array(ratio)))
+
+    center = center.astype(np.int)
+
+    x1 = center[0] - radius
+    x2 = center[0] + radius
+    y1 = center[1] - radius
+    y2 = center[1] + radius
+
+    return (x1, y1, x2-x1, y2-y1)
+
+def fullbody_crop(pts):
+    flags = pts[:,2] > 0.5      #openpose
+    # flags = pts[:,2] > 0.2  #detectron
+    check_id = [11,19,21,22]
+    cnt = sum(flags[check_id])
+
+    if cnt == 0:
+        center = pts[8,:2].astype(np.int)
+        pts = pts[pts[:,2] > 0.5][:,:2]
+        radius = int(1.45*np.sqrt(((center[None,:] - pts)**2).sum(1)).max(0))
+        center[1] += int(0.05*radius)
+    else:
+        pts = pts[pts[:,2] > 0.2]
+        pmax = pts.max(0)
+        pmin = pts.min(0)
+
+        center = (0.5 * (pmax[:2] + pmin[:2])).astype(np.int)
+        radius = int(0.65 * max(pmax[0]-pmin[0], pmax[1]-pmin[1]))
+
+    x1 = center[0] - radius
+    x2 = center[0] + radius
+    y1 = center[1] - radius
+    y2 = center[1] + radius
+
+    return (x1, y1, x2-x1, y2-y1)
+
+
+class EvalWPoseDataset(Dataset):
+    @staticmethod
+    def modify_commandline_options(parser, is_train):
+        return parser
+
+    def __init__(self, opt, projection='orthogonal'):
+        self.opt = opt
+        self.projection_mode = projection
+
+        self.root = self.opt.dataroot
+        self.img_files = sorted([os.path.join(self.root,f) for f in os.listdir(self.root) if f.split('.')[-1] in ['png', 'jpeg', 'jpg', 'PNG', 'JPG', 'JPEG'] and os.path.exists(os.path.join(self.root,f.replace('.%s' % (f.split('.')[-1]), '_keypoints.json')))])
+        self.IMG = os.path.join(self.root)
+
+        self.phase = 'val'
+        self.load_size = self.opt.loadSize
+
+        if self.opt.crop_type == 'face':
+            self.crop_func = face_crop
+        elif self.opt.crop_type == 'upperbody':
+            self.crop_func = upperbody_crop
+        else:
+            self.crop_func = fullbody_crop
+
+        # PIL to tensor
+        self.to_tensor = transforms.Compose([
+            transforms.ToTensor(),
+            transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
+        ])
+
+        # only used in case of multi-person processing
+        self.person_id = 0
+
+    def __len__(self):
+        return len(self.img_files)
+
+    def get_n_person(self, index):
+        joint_path = self.img_files[index].replace('.%s' % (self.img_files[index].split('.')[-1]), '_keypoints.json')
+        # Calib
+        with open(joint_path) as json_file:
+            data = json.load(json_file)
+            return len(data['people'])            
+
+    def get_item(self, index):
+        img_path = self.img_files[index]
+        joint_path = self.img_files[index].replace('.%s' % (self.img_files[index].split('.')[-1]), '_keypoints.json')
+        # Name
+        img_name = os.path.splitext(os.path.basename(img_path))[0]
+        # Calib
+        with open(joint_path) as json_file:
+            data = json.load(json_file)
+            if len(data['people']) == 0:
+                raise IOError('non human found!!')
+            
+            # if True, the person with the largest height will be chosen. 
+            # set to False for multi-person processing
+            if True:
+                selected_data = data['people'][0]
+                height = 0
+                if len(data['people']) != 1:
+                    for i in range(len(data['people'])):
+                        tmp = data['people'][i]
+                        keypoints = np.array(tmp['pose_keypoints_2d']).reshape(-1,3)
+
+                        flags = keypoints[:,2] > 0.5 #openpose
+                        # flags = keypoints[:,2] > 0.2  #detectron
+                        if sum(flags) == 0:
+                            continue
+                        bbox = keypoints[flags]
+                        bbox_max = bbox.max(0)
+                        bbox_min = bbox.min(0)
+
+                        if height < bbox_max[1] - bbox_min[1]:
+                            height = bbox_max[1] - bbox_min[1]
+                            selected_data = tmp
+            else:
+                pid = min(len(data['people'])-1, self.person_id)
+                selected_data = data['people'][pid]
+
+            keypoints = np.array(selected_data['pose_keypoints_2d']).reshape(-1,3)
+
+            flags = keypoints[:,2] > 0.5   #openpose
+            # flags = keypoints[:,2] > 0.2    #detectron
+
+            nflag = flags[0]
+            mflag = flags[1]
+
+            check_id = [2, 5, 15, 16, 17, 18]
+            cnt = sum(flags[check_id])
+            if self.opt.crop_type == 'face' and (not (nflag and cnt > 3)):
+                print('Waring: face should not be backfacing.')
+            if self.opt.crop_type == 'upperbody' and (not (mflag and nflag and cnt > 3)):
+                print('Waring: upperbody should not be backfacing.')
+            if self.opt.crop_type == 'fullbody' and sum(flags) < 15:
+                print('Waring: not sufficient keypoints.')
+
+        im = cv2.imread(img_path, cv2.IMREAD_UNCHANGED)
+        if im.shape[2] == 4:
+            im = im / 255.0
+            im[:,:,:3] /= im[:,:,3:] + 1e-8
+            im = im[:,:,3:] * im[:,:,:3] + 0.5 * (1.0 - im[:,:,3:])
+            im = (255.0 * im).astype(np.uint8)
+        h, w = im.shape[:2]
+        
+        intrinsic = np.identity(4)
+
+        trans_mat = np.identity(4)
+        rect = self.crop_func(keypoints)
+
+        im = crop_image(im, rect)
+
+        scale_im2ndc = 1.0 / float(w // 2)
+        scale = w / rect[2]
+        trans_mat *= scale
+        trans_mat[3,3] = 1.0
+        trans_mat[0, 3] = -scale*(rect[0] + rect[2]//2 - w//2) * scale_im2ndc
+        trans_mat[1, 3] = scale*(rect[1] + rect[3]//2 - h//2) * scale_im2ndc
+        
+        intrinsic = np.matmul(trans_mat, intrinsic)
+        im_512 = cv2.resize(im, (512, 512))
+        im = cv2.resize(im, (self.load_size, self.load_size))
+
+        image_512 = Image.fromarray(im_512[:,:,::-1]).convert('RGB')
+        image = Image.fromarray(im[:,:,::-1]).convert('RGB')
+        
+        B_MIN = np.array([-1, -1, -1])
+        B_MAX = np.array([1, 1, 1])
+        projection_matrix = np.identity(4)
+        projection_matrix[1, 1] = -1
+        calib = torch.Tensor(projection_matrix).float()
+
+        calib_world = torch.Tensor(intrinsic).float()
+
+        # image
+        image_512 = self.to_tensor(image_512)
+        image = self.to_tensor(image)
+        return {
+            'name': img_name,
+            'img': image.unsqueeze(0),
+            'img_512': image_512.unsqueeze(0),
+            'calib': calib.unsqueeze(0),
+            'calib_world': calib_world.unsqueeze(0),
+            'b_min': B_MIN,
+            'b_max': B_MAX,
+        }
+
+    def __getitem__(self, index):
+        return self.get_item(index)

pifuhd/lib/data/__init__.py

@@ -0,0 +1,4 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All rights reserved.
+
+from .EvalWPoseDataset import EvalWPoseDataset
+from .EvalDataset import EvalDataset

pifuhd/lib/evaluator.py

@@ -0,0 +1,233 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All rights reserved.
+
+import trimesh
+import trimesh.proximity
+import trimesh.sample
+import numpy as np
+import math
+import os
+from PIL import Image
+
+import argparse
+
+def euler_to_rot_mat(r_x, r_y, r_z):
+    R_x = np.array([[1, 0, 0],
+                    [0, math.cos(r_x), -math.sin(r_x)],
+                    [0, math.sin(r_x), math.cos(r_x)]
+                    ])
+
+    R_y = np.array([[math.cos(r_y), 0, math.sin(r_y)],
+                    [0, 1, 0],
+                    [-math.sin(r_y), 0, math.cos(r_y)]
+                    ])
+
+    R_z = np.array([[math.cos(r_z), -math.sin(r_z), 0],
+                    [math.sin(r_z), math.cos(r_z), 0],
+                    [0, 0, 1]
+                    ])
+
+    R = np.dot(R_z, np.dot(R_y, R_x))
+
+    return R
+
+
+class MeshEvaluator:
+    _normal_render = None
+
+    @staticmethod
+    def init_gl():
+        from .render.gl.normal_render import NormalRender
+        MeshEvaluator._normal_render = NormalRender(width=512, height=512)
+
+    def __init__(self):
+        pass
+
+    def set_mesh(self, src_path, tgt_path, scale_factor=1.0, offset=0):
+        self.src_mesh = trimesh.load(src_path)
+        self.tgt_mesh = trimesh.load(tgt_path)
+
+        self.scale_factor = scale_factor
+        self.offset = offset
+
+
+    def get_chamfer_dist(self, num_samples=10000):
+        # Chamfer
+        src_surf_pts, _ = trimesh.sample.sample_surface(self.src_mesh, num_samples)
+        tgt_surf_pts, _ = trimesh.sample.sample_surface(self.tgt_mesh, num_samples)
+
+        _, src_tgt_dist, _ = trimesh.proximity.closest_point(self.tgt_mesh, src_surf_pts)
+        _, tgt_src_dist, _ = trimesh.proximity.closest_point(self.src_mesh, tgt_surf_pts)
+
+        src_tgt_dist[np.isnan(src_tgt_dist)] = 0
+        tgt_src_dist[np.isnan(tgt_src_dist)] = 0
+
+        src_tgt_dist = src_tgt_dist.mean()
+        tgt_src_dist = tgt_src_dist.mean()
+
+        chamfer_dist = (src_tgt_dist + tgt_src_dist) / 2
+
+        return chamfer_dist
+
+    def get_surface_dist(self, num_samples=10000):
+        # P2S
+        src_surf_pts, _ = trimesh.sample.sample_surface(self.src_mesh, num_samples)
+
+        _, src_tgt_dist, _ = trimesh.proximity.closest_point(self.tgt_mesh, src_surf_pts)
+
+        src_tgt_dist[np.isnan(src_tgt_dist)] = 0
+
+        src_tgt_dist = src_tgt_dist.mean()
+
+        return src_tgt_dist
+
+    def _render_normal(self, mesh, deg):
+        view_mat = np.identity(4)
+        view_mat[:3, :3] *= 2 / 256
+        rz = deg / 180. * np.pi
+        model_mat = np.identity(4)
+        model_mat[:3, :3] = euler_to_rot_mat(0, rz, 0)
+        model_mat[1, 3] = self.offset
+        view_mat[2, 2] *= -1
+
+        self._normal_render.set_matrices(view_mat, model_mat)
+        self._normal_render.set_normal_mesh(self.scale_factor*mesh.vertices, mesh.faces, mesh.vertex_normals, mesh.faces)
+        self._normal_render.draw()
+        normal_img = self._normal_render.get_color()
+        return normal_img
+
+    def _get_reproj_normal_error(self, deg):
+        tgt_normal = self._render_normal(self.tgt_mesh, deg)
+        src_normal = self._render_normal(self.src_mesh, deg)
+
+        error = ((src_normal[:, :, :3] - tgt_normal[:, :, :3]) ** 2).mean() * 3
+
+        return error, src_normal, tgt_normal
+
+    def get_reproj_normal_error(self, frontal=True, back=True, left=True, right=True, save_demo_img=None):
+        # reproj error
+        # if save_demo_img is not None, save a visualization at the given path (etc, "./test.png")
+        if self._normal_render is None:
+            print("In order to use normal render, "
+                  "you have to call init_gl() before initialing any evaluator objects.")
+            return -1
+
+        side_cnt = 0
+        total_error = 0
+        demo_list = []
+        if frontal:
+            side_cnt += 1
+            error, src_normal, tgt_normal = self._get_reproj_normal_error(0)
+            total_error += error
+            demo_list.append(np.concatenate([src_normal, tgt_normal], axis=0))
+        if back:
+            side_cnt += 1
+            error, src_normal, tgt_normal = self._get_reproj_normal_error(180)
+            total_error += error
+            demo_list.append(np.concatenate([src_normal, tgt_normal], axis=0))
+        if left:
+            side_cnt += 1
+            error, src_normal, tgt_normal = self._get_reproj_normal_error(90)
+            total_error += error
+            demo_list.append(np.concatenate([src_normal, tgt_normal], axis=0))
+        if right:
+            side_cnt += 1
+            error, src_normal, tgt_normal = self._get_reproj_normal_error(270)
+            total_error += error
+            demo_list.append(np.concatenate([src_normal, tgt_normal], axis=0))
+        if save_demo_img is not None:
+            res_array = np.concatenate(demo_list, axis=1)
+            res_img = Image.fromarray((res_array * 255).astype(np.uint8))
+            res_img.save(save_demo_img)
+        return total_error / side_cnt
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument('-r', '--root', type=str, required=True)
+    parser.add_argument('-t', '--tar_path', type=str, required=True)
+    args = parser.parse_args()
+
+    evaluator = MeshEvaluator()
+    evaluator.init_gl()
+
+    def run(root, exp_name, tar_path):
+        src_path = os.path.join(root, exp_name, 'recon')
+        rp_path = os.path.join(tar_path, 'RP', 'GEO', 'OBJ')
+        bf_path = os.path.join(tar_path, 'BUFF', 'GEO', 'PLY')
+
+        buff_files = [f for f in os.listdir(bf_path) if '.ply' in f]
+
+        src_names = ['0_0_00.obj', '90_0_00.obj', '180_0_00.obj', '270_0_00.obj']
+
+        total_vals = []
+        items = []
+        for file in buff_files:
+            tar_name = os.path.join(bf_path, file)
+            name = tar_name.split('/')[-1][:-4]
+
+            for src in src_names:
+                src_name = os.path.join(src_path, 'result_%s_%s' % (name, src))
+                if not os.path.exists(src_name):
+                    continue
+                evaluator.set_mesh(src_name, tar_name, 0.13, -40)
+
+                vals = []
+                vals.append(0.1 * evaluator.get_chamfer_dist())
+                vals.append(0.1 * evaluator.get_surface_dist())
+                vals.append(4.0 * evaluator.get_reproj_normal_error(save_demo_img=os.path.join(src_path, '%s_%s.png' % (name, src[:-4]))))
+
+                item = {
+                    'name': '%s_%s' % (name, src),
+                    'vals': vals
+                }
+
+                total_vals.append(vals)
+                items.append(item)
+
+        vals = np.array(total_vals).mean(0)
+        buf_val = vals
+
+        np.save(os.path.join(root, exp_name, 'buff-item.npy'), np.array(items))
+        np.save(os.path.join(root, exp_name, 'buff-vals.npy'), total_vals)
+
+        rp_files = [f for f in os.listdir(rp_path) if '.obj' in f]
+
+        total_vals = []
+        items = []
+        for file in rp_files:
+            tar_name = os.path.join(rp_path, file)
+            name = tar_name.split('/')[-1][:-9]
+
+            for src in src_names:
+                src_name = os.path.join(src_path, 'result_%s_%s' % (name, src))
+                if not os.path.exists(src_name):
+                    continue
+
+                evaluator.set_mesh(src_name, tar_name, 1.3, -120)
+
+                vals = []
+                vals.append(evaluator.get_chamfer_dist())
+                vals.append(evaluator.get_surface_dist())
+                vals.append(4.0 * evaluator.get_reproj_normal_error(save_demo_img=os.path.join(src_path, '%s_%s.png' % (name, src[:-4]))))
+
+                item = {
+                    'name': '%s_%s' % (name, src),
+                    'vals': vals
+                }
+
+                total_vals.append(vals)
+                items.append(item)
+
+        np.save(os.path.join(root, exp_name, 'rp-item.npy'), np.array(items))
+        np.save(os.path.join(root, exp_name, 'rp-vals.npy'), total_vals)
+
+        vals = np.array(total_vals).mean(0)
+        print('BUFF - chamfer: %.4f  p2s: %.4f  nml: %.4f' % (buf_val[0], buf_val[1], buf_val[2]))
+        print('RP - chamfer: %.4f  p2s: %.4f  nml: %.4f' % (vals[0], vals[1], vals[2]))
+
+    exp_list = ['pifuhd_final']
+
+    root = args.root
+    tar_path = args.tar_path
+
+    for exp in exp_list:
+        run(root, exp, tar_path)

pifuhd/lib/geometry.py

@@ -0,0 +1,79 @@
+'''
+MIT License
+
+Copyright (c) 2019 Shunsuke Saito, Zeng Huang, and Ryota Natsume
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+'''
+import torch
+
+def index(feat, uv):
+    '''
+    extract image features at floating coordinates with bilinear interpolation
+    args:
+        feat: [B, C, H, W] image features
+        uv: [B, 2, N] normalized image coordinates ranged in [-1, 1]
+    return:
+        [B, C, N] sampled pixel values
+    '''
+    uv = uv.transpose(1, 2)
+    uv = uv.unsqueeze(2)
+    samples = torch.nn.functional.grid_sample(feat, uv, align_corners=True)
+    return samples[:, :, :, 0]
+
+def orthogonal(points, calib, transform=None):
+    '''
+    project points onto screen space using orthogonal projection
+    args:
+        points: [B, 3, N] 3d points in world coordinates
+        calib: [B, 3, 4] projection matrix
+        transform: [B, 2, 3] screen space transformation
+    return:
+        [B, 3, N] 3d coordinates in screen space
+    '''
+    rot = calib[:, :3, :3]
+    trans = calib[:, :3, 3:4]
+    pts = torch.baddbmm(trans, rot, points)
+    if transform is not None:
+        scale = transform[:2, :2]
+        shift = transform[:2, 2:3]
+        pts[:, :2, :] = torch.baddbmm(shift, scale, pts[:, :2, :])
+    return pts
+
+def perspective(points, calib, transform=None):
+    '''
+    project points onto screen space using perspective projection
+    args:
+        points: [B, 3, N] 3d points in world coordinates
+        calib: [B, 3, 4] projection matrix
+        transform: [B, 2, 3] screen space trasnformation
+    return:
+        [B, 3, N] 3d coordinates in screen space
+    '''
+    rot = calib[:, :3, :3]
+    trans = calib[:, :3, 3:4]
+    homo = torch.baddbmm(trans, rot, points)
+    xy = homo[:, :2, :] / homo[:, 2:3, :]
+    if transform is not None:
+        scale = transform[:2, :2]
+        shift = transform[:2, 2:3]
+        xy = torch.baddbmm(shift, scale, xy)
+    
+    xyz = torch.cat([xy, homo[:, 2:3, :]], 1)
+    return xyz

pifuhd/lib/mesh_util.py

@@ -0,0 +1,130 @@
+'''
+MIT License
+
+Copyright (c) 2019 Shunsuke Saito, Zeng Huang, and Ryota Natsume
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+'''
+from skimage import measure
+import numpy as np
+import torch
+from .sdf import create_grid, eval_grid_octree, eval_grid
+from skimage import measure
+
+from numpy.linalg import inv
+
+def reconstruction(net, cuda, calib_tensor,
+                   resolution, b_min, b_max, thresh=0.5,
+                   use_octree=False, num_samples=10000, transform=None):
+    '''
+    Reconstruct meshes from sdf predicted by the network.
+    :param net: a BasePixImpNet object. call image filter beforehead.
+    :param cuda: cuda device
+    :param calib_tensor: calibration tensor
+    :param resolution: resolution of the grid cell
+    :param b_min: bounding box corner [x_min, y_min, z_min]
+    :param b_max: bounding box corner [x_max, y_max, z_max]
+    :param use_octree: whether to use octree acceleration
+    :param num_samples: how many points to query each gpu iteration
+    :return: marching cubes results.
+    '''
+    # First we create a grid by resolution
+    # and transforming matrix for grid coordinates to real world xyz
+    coords, mat = create_grid(resolution, resolution, resolution)
+                              #b_min, b_max, transform=transform)
+
+    calib = calib_tensor[0].cpu().numpy()
+
+    calib_inv = inv(calib)
+    coords = coords.reshape(3,-1).T
+    coords = np.matmul(np.concatenate([coords, np.ones((coords.shape[0],1))], 1), calib_inv.T)[:, :3]
+    coords = coords.T.reshape(3,resolution,resolution,resolution)
+
+    # Then we define the lambda function for cell evaluation
+    def eval_func(points):
+        points = np.expand_dims(points, axis=0)
+        points = np.repeat(points, 1, axis=0)
+        samples = torch.from_numpy(points).to(device=cuda).float()
+        
+        net.query(samples, calib_tensor)
+        pred = net.get_preds()[0][0]
+        return pred.detach().cpu().numpy()
+
+    # Then we evaluate the grid
+    if use_octree:
+        sdf = eval_grid_octree(coords, eval_func, num_samples=num_samples)
+    else:
+        sdf = eval_grid(coords, eval_func, num_samples=num_samples)
+
+    # Finally we do marching cubes
+    try:
+        verts, faces, normals, values = measure.marching_cubes_lewiner(sdf, thresh)
+        # transform verts into world coordinate system
+        trans_mat = np.matmul(calib_inv, mat)
+        verts = np.matmul(trans_mat[:3, :3], verts.T) + trans_mat[:3, 3:4]
+        verts = verts.T
+        # in case mesh has flip transformation
+        if np.linalg.det(trans_mat[:3, :3]) < 0.0:
+            faces = faces[:,::-1]
+        return verts, faces, normals, values
+    except:
+        print('error cannot marching cubes')
+        return -1
+
+
+def save_obj_mesh(mesh_path, verts, faces=None):
+    file = open(mesh_path, 'w')
+
+    for v in verts:
+        file.write('v %.4f %.4f %.4f\n' % (v[0], v[1], v[2]))
+    if faces is not None:
+        for f in faces:
+            if f[0] == f[1] or f[1] == f[2] or f[0] == f[2]:
+                continue
+            f_plus = f + 1
+            file.write('f %d %d %d\n' % (f_plus[0], f_plus[2], f_plus[1]))
+    file.close()
+
+
+def save_obj_mesh_with_color(mesh_path, verts, faces, colors):
+    file = open(mesh_path, 'w')
+
+    for idx, v in enumerate(verts):
+        c = colors[idx]
+        file.write('v %.4f %.4f %.4f %.4f %.4f %.4f\n' % (v[0], v[1], v[2], c[0], c[1], c[2]))
+    for f in faces:
+        f_plus = f + 1
+        file.write('f %d %d %d\n' % (f_plus[0], f_plus[2], f_plus[1]))
+    file.close()
+
+
+def save_obj_mesh_with_uv(mesh_path, verts, faces, uvs):
+    file = open(mesh_path, 'w')
+
+    for idx, v in enumerate(verts):
+        vt = uvs[idx]
+        file.write('v %.4f %.4f %.4f\n' % (v[0], v[1], v[2]))
+        file.write('vt %.4f %.4f\n' % (vt[0], vt[1]))
+
+    for f in faces:
+        f_plus = f + 1
+        file.write('f %d/%d %d/%d %d/%d\n' % (f_plus[0], f_plus[0],
+                                              f_plus[2], f_plus[2],
+                                              f_plus[1], f_plus[1]))
+    file.close()

pifuhd/lib/model/BasePIFuNet.py

@@ -0,0 +1,98 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All rights reserved.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F 
+
+from ..geometry import index, orthogonal, perspective
+
+class BasePIFuNet(nn.Module):
+    def __init__(self,
+                 projection_mode='orthogonal',
+                 criteria={'occ': nn.MSELoss()},
+                 ):
+        '''
+        args:
+            projection_mode: orthonal / perspective
+            error_term: point-wise error term 
+        '''
+        super(BasePIFuNet, self).__init__()
+        self.name = 'base'
+
+        self.criteria = criteria
+
+        self.index = index
+        self.projection = orthogonal if projection_mode == 'orthogonal' else perspective
+
+        self.preds = None
+        self.labels = None
+        self.nmls = None
+        self.labels_nml = None
+        self.preds_surface = None # with normal loss only
+
+    def forward(self, points, images, calibs, transforms=None):
+        '''
+        args:
+            points: [B, 3, N] 3d points in world space
+            images: [B, C, H, W] input images
+            calibs: [B, 3, 4] calibration matrices for each image
+            transforms: [B, 2, 3] image space coordinate transforms
+        return:
+            [B, C, N] prediction corresponding to the given points
+        '''
+        self.filter(images)
+        self.query(points, calibs, transforms)
+        return self.get_preds()
+
+    def filter(self, images):
+        '''
+        apply a fully convolutional network to images.
+        the resulting feature will be stored.
+        args:
+            images: [B, C, H, W]
+        '''
+        None
+    
+    def query(self, points, calibs, trasnforms=None, labels=None):
+        '''
+        given 3d points, we obtain 2d projection of these given the camera matrices.
+        filter needs to be called beforehand.
+        the prediction is stored to self.preds
+        args:
+            points: [B, 3, N] 3d points in world space
+            calibs: [B, 3, 4] calibration matrices for each image
+            transforms: [B, 2, 3] image space coordinate transforms
+            labels: [B, C, N] ground truth labels (for supervision only)
+        return:
+            [B, C, N] prediction
+        '''
+        None
+
+    def calc_normal(self, points, calibs, transforms=None, delta=0.1):
+        '''
+        return surface normal in 'model' space.
+        it computes normal only in the last stack.
+        note that the current implementation use forward difference.
+        args:
+            points: [B, 3, N] 3d points in world space
+            calibs: [B, 3, 4] calibration matrices for each image
+            transforms: [B, 2, 3] image space coordinate transforms
+            delta: perturbation for finite difference
+        '''
+        None
+
+    def get_preds(self):
+        '''
+        return the current prediction.
+        return:
+            [B, C, N] prediction
+        '''
+        return self.preds
+
+    def get_error(self, gamma=None):
+        '''
+        return the loss given the ground truth labels and prediction
+        '''
+        return self.error_term(self.preds, self.labels, gamma)
+
+    

pifuhd/lib/model/DepthNormalizer.py

@@ -0,0 +1,20 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All rights reserved.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F 
+
+class DepthNormalizer(nn.Module):
+    def __init__(self, opt):
+        super(DepthNormalizer, self).__init__()
+        self.opt = opt
+
+    def forward(self, xyz, calibs=None, index_feat=None):
+        '''
+        normalize depth value
+        args:
+            xyz: [B, 3, N] depth value
+        '''
+        z_feat = xyz[:,2:3,:] * (self.opt.loadSize // 2) / self.opt.z_size
+
+        return z_feat

pifuhd/lib/model/HGFilters.py

@@ -0,0 +1,216 @@
+'''
+MIT License
+
+Copyright (c) 2019 Shunsuke Saito, Zeng Huang, and Ryota Natsume
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+'''
+import torch
+import torch.nn as nn 
+import torch.nn.functional as F 
+from ..net_util import conv3x3
+
+class ConvBlock(nn.Module):
+    def __init__(self, in_planes, out_planes, norm='batch'):
+        super(ConvBlock, self).__init__()
+        self.conv1 = conv3x3(in_planes, int(out_planes / 2))
+        self.conv2 = conv3x3(int(out_planes / 2), int(out_planes / 4))
+        self.conv3 = conv3x3(int(out_planes / 4), int(out_planes / 4))
+
+        if norm == 'batch':
+            self.bn1 = nn.BatchNorm2d(in_planes)
+            self.bn2 = nn.BatchNorm2d(int(out_planes / 2))
+            self.bn3 = nn.BatchNorm2d(int(out_planes / 4))
+            self.bn4 = nn.BatchNorm2d(in_planes)
+        elif norm == 'group':
+            self.bn1 = nn.GroupNorm(32, in_planes)
+            self.bn2 = nn.GroupNorm(32, int(out_planes / 2))
+            self.bn3 = nn.GroupNorm(32, int(out_planes / 4))
+            self.bn4 = nn.GroupNorm(32, in_planes)
+        
+        if in_planes != out_planes:
+            self.downsample = nn.Sequential(
+                self.bn4,
+                nn.ReLU(True),
+                nn.Conv2d(in_planes, out_planes,
+                          kernel_size=1, stride=1, bias=False),
+            )
+        else:
+            self.downsample = None
+    
+    def forward(self, x):
+        residual = x
+
+        out1 = self.conv1(F.relu(self.bn1(x), True))
+        out2 = self.conv2(F.relu(self.bn2(out1), True))
+        out3 = self.conv3(F.relu(self.bn3(out2), True))
+
+        out3 = torch.cat([out1, out2, out3], 1)
+
+        if self.downsample is not None:
+            residual = self.downsample(residual)
+        
+        out3 += residual
+
+        return out3
+
+class HourGlass(nn.Module):
+    def __init__(self, depth, n_features, norm='batch'):
+        super(HourGlass, self).__init__()
+        self.depth = depth
+        self.features = n_features
+        self.norm = norm
+
+        self._generate_network(self.depth)
+
+    def _generate_network(self, level):
+        self.add_module('b1_' + str(level), ConvBlock(self.features, self.features, norm=self.norm))
+        self.add_module('b2_' + str(level), ConvBlock(self.features, self.features, norm=self.norm))
+
+        if level > 1:
+            self._generate_network(level - 1)
+        else:
+            self.add_module('b2_plus_' + str(level), ConvBlock(self.features, self.features, norm=self.norm))
+
+        self.add_module('b3_' + str(level), ConvBlock(self.features, self.features, norm=self.norm))
+
+    def _forward(self, level, inp):
+        # upper branch
+        up1 = inp 
+        up1 = self._modules['b1_' + str(level)](up1)
+
+        # lower branch
+        low1 = F.avg_pool2d(inp, 2, stride=2)
+        low1 = self._modules['b2_' + str(level)](low1)
+
+        if level > 1:
+            low2 = self._forward(level - 1, low1)
+        else:
+            low2 = low1
+            low2 = self._modules['b2_plus_' + str(level)](low2)
+
+        low3 = low2
+        low3 = self._modules['b3_' + str(level)](low3)
+
+        up2 = F.interpolate(low3, scale_factor=2, mode='bicubic', align_corners=True)
+        # up2 = F.interpolate(low3, scale_factor=2, mode='bilinear')
+
+        return up1 + up2
+    
+    def forward(self, x):
+        return self._forward(self.depth, x)
+        
+
+class HGFilter(nn.Module):
+    def __init__(self, stack, depth, in_ch, last_ch, norm='batch', down_type='conv64', use_sigmoid=True):
+        super(HGFilter, self).__init__()
+        self.n_stack = stack
+        self.use_sigmoid = use_sigmoid
+        self.depth = depth
+        self.last_ch = last_ch
+        self.norm = norm
+        self.down_type = down_type
+
+        self.conv1 = nn.Conv2d(in_ch, 64, kernel_size=7, stride=2, padding=3)
+
+        last_ch = self.last_ch
+
+        if self.norm == 'batch':
+            self.bn1 = nn.BatchNorm2d(64)
+        elif self.norm == 'group':
+            self.bn1 = nn.GroupNorm(32, 64)
+
+        if self.down_type == 'conv64':
+            self.conv2 = ConvBlock(64, 64, self.norm)
+            self.down_conv2 = nn.Conv2d(64, 128, kernel_size=3, stride=2, padding=1)
+        elif self.down_type == 'conv128':
+            self.conv2 = ConvBlock(128, 128, self.norm)
+            self.down_conv2 = nn.Conv2d(128, 128, kernel_size=3, stride=2, padding=1)
+        elif self.down_type == 'ave_pool' or self.down_type == 'no_down':
+            self.conv2 = ConvBlock(64, 128, self.norm)
+        
+        self.conv3 = ConvBlock(128, 128, self.norm)
+        self.conv4 = ConvBlock(128, 256, self.norm)
+        
+        # start stacking
+        for stack in range(self.n_stack):
+            self.add_module('m' + str(stack), HourGlass(self.depth, 256, self.norm))
+
+            self.add_module('top_m_' + str(stack), ConvBlock(256, 256, self.norm))
+            self.add_module('conv_last' + str(stack),
+                            nn.Conv2d(256, 256, kernel_size=1, stride=1, padding=0))
+            if self.norm == 'batch':
+                self.add_module('bn_end' + str(stack), nn.BatchNorm2d(256))
+            elif self.norm == 'group':
+                self.add_module('bn_end' + str(stack), nn.GroupNorm(32, 256))
+            
+            self.add_module('l' + str(stack),
+                            nn.Conv2d(256, last_ch, 
+                            kernel_size=1, stride=1, padding=0))
+            
+            if stack < self.n_stack - 1:
+                self.add_module(
+                    'bl' + str(stack), nn.Conv2d(256, 256, kernel_size=1, stride=1, padding=0))
+                self.add_module(
+                    'al' + str(stack), nn.Conv2d(last_ch, 256, kernel_size=1, stride=1, padding=0))
+
+    def forward(self, x):
+        x = F.relu(self.bn1(self.conv1(x)), True)
+
+        if self.down_type == 'ave_pool':
+            x = F.avg_pool2d(self.conv2(x), 2, stride=2)
+        elif self.down_type == ['conv64', 'conv128']:
+            x = self.conv2(x)
+            x = self.down_conv2(x)
+        elif self.down_type == 'no_down':
+            x = self.conv2(x)
+        else:
+            raise NameError('unknown downsampling type')
+    
+        normx = x
+
+        x = self.conv3(x)
+        x = self.conv4(x)
+
+        previous = x
+
+        outputs = []
+        for i in range(self.n_stack):
+            hg = self._modules['m' + str(i)](previous)
+
+            ll = hg
+            ll = self._modules['top_m_' + str(i)](ll)
+
+            ll = F.relu(self._modules['bn_end' + str(i)]
+                       (self._modules['conv_last' + str(i)](ll)), True)
+
+            tmp_out = self._modules['l' + str(i)](ll)
+
+            if self.use_sigmoid:
+                outputs.append(nn.Tanh()(tmp_out))
+            else:
+                outputs.append(tmp_out)
+            
+            if i < self.n_stack - 1:
+                ll = self._modules['bl' + str(i)](ll)
+                tmp_out_ = self._modules['al' + str(i)](tmp_out)
+                previous = previous + ll + tmp_out_
+            
+        return outputs, normx
+    

pifuhd/lib/model/HGPIFuMRNet.py

@@ -0,0 +1,302 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All rights reserved.
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F 
+from .BasePIFuNet import BasePIFuNet
+from .MLP import MLP
+from .DepthNormalizer import DepthNormalizer
+from .HGFilters import HGFilter
+from ..net_util import init_net
+import cv2
+
+class HGPIFuMRNet(BasePIFuNet):
+    '''
+    HGPIFu uses stacked hourglass as an image encoder.
+    '''
+
+    def __init__(self, 
+                 opt, 
+                 netG,
+                 projection_mode='orthogonal',
+                 criteria={'occ': nn.MSELoss()}
+                 ):
+        super(HGPIFuMRNet, self).__init__(
+            projection_mode=projection_mode,
+            criteria=criteria)
+
+        self.name = 'hg_pifu'
+
+        in_ch = 3
+        try:
+            if netG.opt.use_front_normal:
+                in_ch += 3
+            if netG.opt.use_back_normal:
+                in_ch += 3
+        except:
+            pass
+
+        self.opt = opt
+        self.image_filter = HGFilter(opt.num_stack, opt.hg_depth, in_ch, opt.hg_dim, 
+                                     opt.norm, 'no_down', False)
+
+        self.mlp = MLP(
+            filter_channels=self.opt.mlp_dim,
+            merge_layer=-1,
+            res_layers=self.opt.mlp_res_layers,
+            norm=self.opt.mlp_norm,
+            last_op=nn.Sigmoid())
+
+        self.im_feat_list = []
+        self.preds_interm = None
+        self.preds_low = None
+        self.w = None
+        self.gamma = None
+
+        self.intermediate_preds_list = []
+
+        init_net(self)
+
+        self.netG = netG
+
+    def train(self, mode=True):
+        r"""Sets the module in training mode."""      
+        self.training = mode
+        for module in self.children():
+            module.train(mode)
+        if not self.opt.train_full_pifu:
+            self.netG.eval()
+        return self
+
+    def filter_global(self, images):
+        '''
+        apply a fully convolutional network to images.
+        the resulting feature will be stored.
+        args:
+            images: [B1, C, H, W]
+        '''
+        if self.opt.train_full_pifu:
+            self.netG.filter(images)
+        else:
+            with torch.no_grad():
+                self.netG.filter(images)
+
+    def filter_local(self, images, rect=None):
+        '''
+        apply a fully convolutional network to images.
+        the resulting feature will be stored.
+        args:
+            images: [B1, B2, C, H, W]
+        '''
+        nmls = []
+        try:
+            if self.netG.opt.use_front_normal:
+                nmls.append(self.netG.nmlF)
+            if self.netG.opt.use_back_normal:
+                nmls.append(self.netG.nmlB)
+        except:
+            pass
+
+        if len(nmls):
+            nmls = nn.Upsample(size=(self.opt.loadSizeBig,self.opt.loadSizeBig), mode='bilinear', align_corners=True)(torch.cat(nmls,1))
+            
+            # it's kind of damn way.
+            if rect is None:
+                images = torch.cat([images, nmls[:,None].expand(-1,images.size(1),-1,-1,-1)], 2)
+            else:
+                nml = []
+                for i in range(rect.size(0)):
+                    for j in range(rect.size(1)):
+                        x1, y1, x2, y2 = rect[i,j]
+                        tmp = nmls[i,:,y1:y2,x1:x2]
+                        nml.append(nmls[i,:,y1:y2,x1:x2])
+                nml = torch.stack(nml, 0).view(*rect.shape[:2],*nml[0].size())
+                images = torch.cat([images, nml], 2)
+
+        self.im_feat_list, self.normx = self.image_filter(images.view(-1,*images.size()[2:]))
+        if not self.training:
+            self.im_feat_list = [self.im_feat_list[-1]]
+        
+    def query(self, points, calib_local, calib_global=None, transforms=None, labels=None):
+        '''
+        given 3d points, we obtain 2d projection of these given the camera matrices.
+        filter needs to be called beforehand.
+        the prediction is stored to self.preds
+        args:
+            points: [B1, B2, 3, N] 3d points in world space
+            calibs_local: [B1, B2, 4, 4] calibration matrices for each image
+            calibs_global: [B1, 4, 4] calibration matrices for each image
+            transforms: [B1, 2, 3] image space coordinate transforms
+            labels: [B1, B2, C, N] ground truth labels (for supervision only)
+        return:
+            [B, C, N] prediction
+        '''
+        if calib_global is not None:
+            B = calib_local.size(1)
+        else:
+            B = 1
+            points = points[:,None]
+            calib_global = calib_local
+            calib_local = calib_local[:,None]
+
+        ws = []
+        preds = []
+        preds_interm = []
+        preds_low = []
+        gammas = []
+        newlabels = []
+        for i in range(B):
+            xyz = self.projection(points[:,i], calib_local[:,i], transforms)
+            
+            xy = xyz[:, :2, :]
+
+            # if the point is outside bounding box, return outside.
+            in_bb = (xyz >= -1) & (xyz <= 1)
+            in_bb = in_bb[:, 0, :] & in_bb[:, 1, :]
+            in_bb = in_bb[:, None, :].detach().float()
+
+            self.netG.query(points=points[:,i], calibs=calib_global)
+            preds_low.append(torch.stack(self.netG.intermediate_preds_list,0))
+
+            if labels is not None:
+                newlabels.append(in_bb * labels[:,i])
+                with torch.no_grad():
+                    ws.append(in_bb.size(2) / in_bb.view(in_bb.size(0),-1).sum(1))
+                    gammas.append(1 - newlabels[-1].view(newlabels[-1].size(0),-1).sum(1) / in_bb.view(in_bb.size(0),-1).sum(1))
+
+            z_feat = self.netG.phi
+            if not self.opt.train_full_pifu:
+                z_feat = z_feat.detach()
+                        
+            intermediate_preds_list = []
+            for j, im_feat in enumerate(self.im_feat_list):
+                point_local_feat_list = [self.index(im_feat.view(-1,B,*im_feat.size()[1:])[:,i], xy), z_feat]
+                point_local_feat = torch.cat(point_local_feat_list, 1)
+                pred = self.mlp(point_local_feat)[0]
+                pred = in_bb * pred
+                intermediate_preds_list.append(pred)
+
+            preds_interm.append(torch.stack(intermediate_preds_list,0))
+            preds.append(intermediate_preds_list[-1])
+
+        self.preds = torch.cat(preds,0)
+        self.preds_interm = torch.cat(preds_interm, 1) # first dim is for intermediate predictions
+        self.preds_low = torch.cat(preds_low, 1) # first dim is for intermediate predictions
+        
+        if labels is not None:
+            self.w = torch.cat(ws,0)
+            self.gamma = torch.cat(gammas,0)
+            self.labels = torch.cat(newlabels,0)
+
+    def calc_normal(self, points, calib_local, calib_global, transforms=None, labels=None, delta=0.001, fd_type='forward'):
+        '''
+        return surface normal in 'model' space.
+        it computes normal only in the last stack.
+        note that the current implementation use forward difference.
+        args:
+            points: [B1, B2, 3, N] 3d points in world space
+            calibs_local: [B1, B2, 4, 4] calibration matrices for each image
+            calibs_global: [B1, 4, 4] calibration matrices for each image
+            transforms: [B1, 2, 3] image space coordinate transforms
+            labels: [B1, B2, 3, N] ground truth normal
+            delta: perturbation for finite difference
+            fd_type: finite difference type (forward/backward/central) 
+        '''
+        B = calib_local.size(1)
+
+        if labels is not None:
+            self.labels_nml = labels.view(-1,*labels.size()[2:])
+
+        im_feat = self.im_feat_list[-1].view(-1,B,*self.im_feat_list[-1].size()[1:])
+
+        nmls = []
+        for i in range(B):
+            points_sub = points[:,i]
+            pdx = points_sub.clone()
+            pdx[:,0,:] += delta
+            pdy = points_sub.clone()
+            pdy[:,1,:] += delta
+            pdz = points_sub.clone()
+            pdz[:,2,:] += delta
+
+            points_all = torch.stack([points_sub, pdx, pdy, pdz], 3)
+            points_all = points_all.view(*points_sub.size()[:2],-1)
+            xyz = self.projection(points_all, calib_local[:,i], transforms)
+            xy = xyz[:, :2, :]
+
+
+            self.netG.query(points=points_all, calibs=calib_global, update_pred=False)
+            z_feat = self.netG.phi
+            if not self.opt.train_full_pifu:
+                z_feat = z_feat.detach()
+
+            point_local_feat_list = [self.index(im_feat[:,i], xy), z_feat]            
+            point_local_feat = torch.cat(point_local_feat_list, 1)
+            pred = self.mlp(point_local_feat)[0]
+
+            pred = pred.view(*pred.size()[:2],-1,4) # (B, 1, N, 4)
+
+            # divide by delta is omitted since it's normalized anyway
+            dfdx = pred[:,:,:,1] - pred[:,:,:,0]
+            dfdy = pred[:,:,:,2] - pred[:,:,:,0]
+            dfdz = pred[:,:,:,3] - pred[:,:,:,0]
+
+            nml = -torch.cat([dfdx,dfdy,dfdz], 1)
+            nml = F.normalize(nml, dim=1, eps=1e-8)
+
+            nmls.append(nml)
+        
+        self.nmls = torch.stack(nmls,1).view(-1,3,points.size(3))
+
+    def get_im_feat(self):
+        '''
+        return the image filter in the last stack
+        return:
+            [B, C, H, W]
+        '''
+        return self.im_feat_list[-1]
+
+    def get_error(self):
+        '''
+        return the loss given the ground truth labels and prediction
+        '''
+
+        error = {}
+        if self.opt.train_full_pifu:
+            if not self.opt.no_intermediate_loss:
+                error['Err(occ)'] = 0.0
+                for i in range(self.preds_low.size(0)):
+                    error['Err(occ)'] += self.criteria['occ'](self.preds_low[i], self.labels, self.gamma, self.w)
+                error['Err(occ)'] /= self.preds_low.size(0)
+
+            error['Err(occ:fine)'] = 0.0
+            for i in range(self.preds_interm.size(0)):
+                error['Err(occ:fine)'] += self.criteria['occ'](self.preds_interm[i], self.labels, self.gamma, self.w)
+            error['Err(occ:fine)'] /= self.preds_interm.size(0)
+
+            if self.nmls is not None and self.labels_nml is not None:
+                error['Err(nml:fine)'] = self.criteria['nml'](self.nmls, self.labels_nml)
+        else:
+            error['Err(occ:fine)'] = 0.0
+            for i in range(self.preds_interm.size(0)):
+                error['Err(occ:fine)'] += self.criteria['occ'](self.preds_interm[i], self.labels, self.gamma, self.w)
+            error['Err(occ:fine)'] /= self.preds_interm.size(0)
+
+            if self.nmls is not None and self.labels_nml is not None:
+                error['Err(nml:fine)'] = self.criteria['nml'](self.nmls, self.labels_nml)
+        
+        return error
+
+
+    def forward(self, images_local, images_global, points, calib_local, calib_global, labels, points_nml=None, labels_nml=None, rect=None):
+        self.filter_global(images_global)
+        self.filter_local(images_local, rect)
+        self.query(points, calib_local, calib_global, labels=labels)
+        if points_nml is not None and labels_nml is not None:
+            self.calc_normal(points_nml, calib_local, calib_global, labels=labels_nml)
+        res = self.get_preds()
+            
+        err = self.get_error()
+
+        return err, res

pifuhd/lib/model/HGPIFuNetwNML.py

@@ -0,0 +1,264 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All rights reserved.
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F 
+from .BasePIFuNet import BasePIFuNet
+from .MLP import MLP
+from .DepthNormalizer import DepthNormalizer
+from .HGFilters import HGFilter
+from ..net_util import init_net
+from ..networks import define_G
+import cv2
+
+class HGPIFuNetwNML(BasePIFuNet):
+    '''
+    HGPIFu uses stacked hourglass as an image encoder.
+    '''
+
+    def __init__(self, 
+                 opt, 
+                 projection_mode='orthogonal',
+                 criteria={'occ': nn.MSELoss()}
+                 ):
+        super(HGPIFuNetwNML, self).__init__(
+            projection_mode=projection_mode,
+            criteria=criteria)
+
+        self.name = 'hg_pifu'
+
+        in_ch = 3
+        try:
+            if opt.use_front_normal:
+                in_ch += 3
+            if opt.use_back_normal:
+                in_ch += 3
+        except:
+            pass
+        self.opt = opt
+        self.image_filter = HGFilter(opt.num_stack, opt.hg_depth, in_ch, opt.hg_dim, 
+                                     opt.norm, opt.hg_down, False)
+
+        self.mlp = MLP(
+            filter_channels=self.opt.mlp_dim,
+            merge_layer=self.opt.merge_layer,
+            res_layers=self.opt.mlp_res_layers,
+            norm=self.opt.mlp_norm,
+            last_op=nn.Sigmoid())
+
+        self.spatial_enc = DepthNormalizer(opt)
+
+        self.im_feat_list = []
+        self.tmpx = None
+        self.normx = None
+        self.phi = None
+
+        self.intermediate_preds_list = []
+
+        init_net(self)
+
+        self.netF = None
+        self.netB = None
+        try:
+            if opt.use_front_normal:
+                self.netF = define_G(3, 3, 64, "global", 4, 9, 1, 3, "instance")
+            if opt.use_back_normal:
+                self.netB = define_G(3, 3, 64, "global", 4, 9, 1, 3, "instance")
+        except:
+            pass
+        self.nmlF = None
+        self.nmlB = None
+
+    def loadFromHGHPIFu(self, net):
+        hgnet = net.image_filter
+        pretrained_dict = hgnet.state_dict()            
+        model_dict = self.image_filter.state_dict()
+
+        pretrained_dict = {k: v for k, v in hgnet.state_dict().items() if k in model_dict}                    
+
+        for k, v in pretrained_dict.items():                      
+            if v.size() == model_dict[k].size():
+                model_dict[k] = v
+
+        not_initialized = set()
+               
+        for k, v in model_dict.items():
+            if k not in pretrained_dict or v.size() != pretrained_dict[k].size():
+                not_initialized.add(k.split('.')[0])
+        
+        print('not initialized', sorted(not_initialized))
+        self.image_filter.load_state_dict(model_dict) 
+
+        pretrained_dict = net.mlp.state_dict()            
+        model_dict = self.mlp.state_dict()
+
+        pretrained_dict = {k: v for k, v in net.mlp.state_dict().items() if k in model_dict}                    
+
+        for k, v in pretrained_dict.items():                      
+            if v.size() == model_dict[k].size():
+                model_dict[k] = v
+
+        not_initialized = set()
+               
+        for k, v in model_dict.items():
+            if k not in pretrained_dict or v.size() != pretrained_dict[k].size():
+                not_initialized.add(k.split('.')[0])
+        
+        print('not initialized', sorted(not_initialized))
+        self.mlp.load_state_dict(model_dict) 
+
+    def filter(self, images):
+        '''
+        apply a fully convolutional network to images.
+        the resulting feature will be stored.
+        args:
+            images: [B, C, H, W]
+        '''
+        nmls = []
+        # if you wish to train jointly, remove detach etc.
+        with torch.no_grad():
+            if self.netF is not None:
+                self.nmlF = self.netF.forward(images).detach()
+                nmls.append(self.nmlF)
+            if self.netB is not None:
+                self.nmlB = self.netB.forward(images).detach()
+                nmls.append(self.nmlB)
+        if len(nmls) != 0:
+            nmls = torch.cat(nmls,1)
+            if images.size()[2:] != nmls.size()[2:]:
+                nmls = nn.Upsample(size=images.size()[2:], mode='bilinear', align_corners=True)(nmls)
+            images = torch.cat([images,nmls],1)
+
+
+        self.im_feat_list, self.normx = self.image_filter(images)
+
+        if not self.training:
+            self.im_feat_list = [self.im_feat_list[-1]]
+        
+    def query(self, points, calibs, transforms=None, labels=None, update_pred=True, update_phi=True):
+        '''
+        given 3d points, we obtain 2d projection of these given the camera matrices.
+        filter needs to be called beforehand.
+        the prediction is stored to self.preds
+        args:
+            points: [B, 3, N] 3d points in world space
+            calibs: [B, 3, 4] calibration matrices for each image
+            transforms: [B, 2, 3] image space coordinate transforms
+            labels: [B, C, N] ground truth labels (for supervision only)
+        return:
+            [B, C, N] prediction
+        '''
+        xyz = self.projection(points, calibs, transforms)
+        xy = xyz[:, :2, :]
+
+        # if the point is outside bounding box, return outside.
+        in_bb = (xyz >= -1) & (xyz <= 1)
+        in_bb = in_bb[:, 0, :] & in_bb[:, 1, :] & in_bb[:, 2, :]
+        in_bb = in_bb[:, None, :].detach().float()
+
+        if labels is not None:
+            self.labels = in_bb * labels
+
+        sp_feat = self.spatial_enc(xyz, calibs=calibs)
+
+        intermediate_preds_list = []
+
+        phi = None
+        for i, im_feat in enumerate(self.im_feat_list):
+            point_local_feat_list = [self.index(im_feat, xy), sp_feat]       
+            point_local_feat = torch.cat(point_local_feat_list, 1)
+            pred, phi = self.mlp(point_local_feat)
+            pred = in_bb * pred
+
+            intermediate_preds_list.append(pred)
+        
+        if update_phi:
+            self.phi = phi
+
+        if update_pred:
+            self.intermediate_preds_list = intermediate_preds_list
+            self.preds = self.intermediate_preds_list[-1]
+
+    def calc_normal(self, points, calibs, transforms=None, labels=None, delta=0.01, fd_type='forward'):
+        '''
+        return surface normal in 'model' space.
+        it computes normal only in the last stack.
+        note that the current implementation use forward difference.
+        args:
+            points: [B, 3, N] 3d points in world space
+            calibs: [B, 3, 4] calibration matrices for each image
+            transforms: [B, 2, 3] image space coordinate transforms
+            delta: perturbation for finite difference
+            fd_type: finite difference type (forward/backward/central) 
+        '''
+        pdx = points.clone()
+        pdx[:,0,:] += delta
+        pdy = points.clone()
+        pdy[:,1,:] += delta
+        pdz = points.clone()
+        pdz[:,2,:] += delta
+
+        if labels is not None:
+            self.labels_nml = labels
+
+        points_all = torch.stack([points, pdx, pdy, pdz], 3)
+        points_all = points_all.view(*points.size()[:2],-1)
+        xyz = self.projection(points_all, calibs, transforms)
+        xy = xyz[:, :2, :]
+
+        im_feat = self.im_feat_list[-1]
+        sp_feat = self.spatial_enc(xyz, calibs=calibs)
+
+        point_local_feat_list = [self.index(im_feat, xy), sp_feat]            
+        point_local_feat = torch.cat(point_local_feat_list, 1)
+
+        pred = self.mlp(point_local_feat)[0]
+
+        pred = pred.view(*pred.size()[:2],-1,4) # (B, 1, N, 4)
+
+        # divide by delta is omitted since it's normalized anyway
+        dfdx = pred[:,:,:,1] - pred[:,:,:,0]
+        dfdy = pred[:,:,:,2] - pred[:,:,:,0]
+        dfdz = pred[:,:,:,3] - pred[:,:,:,0]
+
+        nml = -torch.cat([dfdx,dfdy,dfdz], 1)
+        nml = F.normalize(nml, dim=1, eps=1e-8)
+
+        self.nmls = nml
+
+    def get_im_feat(self):
+        '''
+        return the image filter in the last stack
+        return:
+            [B, C, H, W]
+        '''
+        return self.im_feat_list[-1]
+
+
+    def get_error(self, gamma):
+        '''
+        return the loss given the ground truth labels and prediction
+        '''
+        error = {}
+        error['Err(occ)'] = 0
+        for preds in self.intermediate_preds_list:
+            error['Err(occ)'] += self.criteria['occ'](preds, self.labels, gamma)
+        
+        error['Err(occ)'] /= len(self.intermediate_preds_list)
+        
+        if self.nmls is not None and self.labels_nml is not None:
+            error['Err(nml)'] = self.criteria['nml'](self.nmls, self.labels_nml)
+
+        return error
+
+    def forward(self, images, points, calibs, labels, gamma, points_nml=None, labels_nml=None):
+        self.filter(images)
+        self.query(points, calibs, labels=labels)
+        if points_nml is not None and labels_nml is not None:
+            self.calc_normal(points_nml, calibs, labels=labels_nml)
+        res = self.get_preds()
+            
+        err = self.get_error(gamma)
+
+        return err, res

pifuhd/lib/model/MLP.py

@@ -0,0 +1,67 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All rights reserved.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F 
+
+class MLP(nn.Module):
+    def __init__(self, 
+                 filter_channels, 
+                 merge_layer=0,
+                 res_layers=[],
+                 norm='group',
+                 last_op=None):
+        super(MLP, self).__init__()
+
+        self.filters = nn.ModuleList()
+        self.norms = nn.ModuleList()
+        self.merge_layer = merge_layer if merge_layer > 0 else len(filter_channels) // 2
+        self.res_layers = res_layers
+        self.norm = norm
+        self.last_op = last_op
+
+        for l in range(0, len(filter_channels)-1):
+            if l in self.res_layers:
+                self.filters.append(nn.Conv1d(
+                    filter_channels[l] + filter_channels[0],
+                    filter_channels[l+1],
+                    1))
+            else:
+                self.filters.append(nn.Conv1d(
+                    filter_channels[l],
+                    filter_channels[l+1],
+                    1))
+            if l != len(filter_channels)-2:
+                if norm == 'group':
+                    self.norms.append(nn.GroupNorm(32, filter_channels[l+1]))
+                elif norm == 'batch':
+                    self.norms.append(nn.BatchNorm1d(filter_channels[l+1]))
+
+    def forward(self, feature):
+        '''
+        feature may include multiple view inputs
+        args:
+            feature: [B, C_in, N]
+        return:
+            [B, C_out, N] prediction
+        '''
+        y = feature
+        tmpy = feature
+        phi = None
+        for i, f in enumerate(self.filters):
+            y = f(
+                y if i not in self.res_layers
+                else torch.cat([y, tmpy], 1)
+            )
+            if i != len(self.filters)-1:
+                if self.norm not in ['batch', 'group']:
+                    y = F.leaky_relu(y)
+                else:
+                    y = F.leaky_relu(self.norms[i](y))         
+            if i == self.merge_layer:
+                phi = y.clone()
+
+        if self.last_op is not None:
+            y = self.last_op(y)
+
+        return y, phi

pifuhd/lib/model/__init__.py

@@ -0,0 +1,5 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All rights reserved.
+
+from .MLP import MLP
+from .HGPIFuMRNet import HGPIFuMRNet
+from .HGPIFuNetwNML import HGPIFuNetwNML

pifuhd/lib/net_util.py

@@ -0,0 +1,159 @@
+'''
+MIT License
+
+Copyright (c) 2019 Shunsuke Saito, Zeng Huang, and Ryota Natsume
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+'''
+import torch
+from torch.nn import init
+import torch.nn as nn
+import torch.nn.functional as F 
+import functools
+
+def load_state_dict(state_dict, net):
+    model_dict = net.state_dict()
+
+    pretrained_dict = {k: v for k, v in state_dict.items() if k in model_dict}                    
+
+    for k, v in pretrained_dict.items():                      
+        if v.size() == model_dict[k].size():
+            model_dict[k] = v
+
+    not_initialized = set()
+            
+    for k, v in model_dict.items():
+        if k not in pretrained_dict or v.size() != pretrained_dict[k].size():
+            not_initialized.add(k.split('.')[0])
+    
+    print('not initialized', sorted(not_initialized))
+    net.load_state_dict(model_dict) 
+
+    return net
+    
+def conv3x3(in_planes, out_planes, strd=1, padding=1, bias=False):
+    return nn.Conv2d(in_planes, out_planes, kernel_size=3,
+                     stride=strd, padding=padding, bias=bias)
+
+def init_weights(net, init_type='normal', init_gain=0.02):
+    def init_func(m):  # define the initialization function
+        classname = m.__class__.__name__
+        if hasattr(m, 'weight') and (classname.find('Conv') != -1 or classname.find('Linear') != -1):
+            if init_type == 'normal':
+                init.normal_(m.weight.data, 0.0, init_gain)
+            elif init_type == 'xavier':
+                init.xavier_normal_(m.weight.data, gain=init_gain)
+            elif init_type == 'kaiming':
+                init.kaiming_normal_(m.weight.data, a=0, mode='fan_in')
+            elif init_type == 'orthogonal':
+                init.orthogonal_(m.weight.data, gain=init_gain)
+            else:
+                raise NotImplementedError('initialization method [%s] is not implemented' % init_type)
+            if hasattr(m, 'bias') and m.bias is not None:
+                init.constant_(m.bias.data, 0.0)
+        elif classname.find(
+                'BatchNorm2d') != -1:
+            init.normal_(m.weight.data, 1.0, init_gain)
+            init.constant_(m.bias.data, 0.0)
+
+    print('initialize network with %s' % init_type)
+    net.apply(init_func)
+
+def init_net(net, init_type='normal', init_gain=0.02, gpu_ids=[]):
+    if len(gpu_ids) > 0:
+        assert (torch.cuda.is_available())
+        net.to(gpu_ids[0])
+        net = torch.nn.DataParallel(net, gpu_ids)  # multi-GPUs
+    init_weights(net, init_type, init_gain=init_gain)
+    return net
+
+class CustomBCELoss(nn.Module):
+    def __init__(self, brock=False, gamma=None):
+        super(CustomBCELoss, self).__init__()
+        self.brock = brock
+        self.gamma = gamma
+
+    def forward(self, pred, gt, gamma, w=None):
+        x_hat = torch.clamp(pred, 1e-5, 1.0-1e-5) # prevent log(0) from happening
+        gamma = gamma[:,None,None] if self.gamma is None else self.gamma
+        if self.brock:
+            x = 3.0*gt - 1.0 # rescaled to [-1,2]
+
+            loss = -(gamma*x*torch.log(x_hat) + (1.0-gamma)*(1.0-x)*torch.log(1.0-x_hat))
+        else:
+            loss = -(gamma*gt*torch.log(x_hat) + (1.0-gamma)*(1.0-gt)*torch.log(1.0-x_hat))
+
+        if w is not None:
+            if len(w.size()) == 1:
+                w = w[:,None,None] 
+            return (loss * w).mean()
+        else:
+            return loss.mean()
+
+class CustomMSELoss(nn.Module):
+    def __init__(self, gamma=None):
+        super(CustomMSELoss, self).__init__()
+        self.gamma = gamma
+
+    def forward(self, pred, gt, gamma, w=None):
+        gamma = gamma[:,None,None] if self.gamma is None else self.gamma
+        weight = gamma * gt + (1.0-gamma) * (1 - gt)
+        loss = (weight * (pred - gt).pow(2)).mean()
+
+        if w is not None:
+            return (loss * w).mean()
+        else:
+            return loss.mean()
+
+def createMLP(dims, norm='bn', activation='relu', last_op=nn.Tanh(), dropout=False):
+    act = None
+    if activation == 'relu':
+        act = nn.ReLU()
+    if activation == 'lrelu':
+        act = nn.LeakyReLU()
+    if activation == 'selu':
+        act = nn.SELU()
+    if activation == 'elu':
+        act = nn.ELU()
+    if activation == 'prelu':
+        act = nn.PReLU()
+
+    mlp = []
+    for i in range(1,len(dims)):
+        if norm == 'bn':
+            mlp += [  nn.Linear(dims[i-1], dims[i]),
+                    nn.BatchNorm1d(dims[i])]
+        if norm == 'in':
+            mlp += [  nn.Linear(dims[i-1], dims[i]),
+                    nn.InstanceNorm1d(dims[i])]
+        if norm == 'wn':
+            mlp += [  nn.utils.weight_norm(nn.Linear(dims[i-1], dims[i]), name='weight')]
+        if norm == 'none':
+            mlp += [ nn.Linear(dims[i-1], dims[i])]
+        
+        if i != len(dims)-1:
+            if act is not None:
+                mlp += [act]
+            if dropout:
+                mlp += [nn.Dropout(0.2)]
+
+    if last_op is not None:
+        mlp += [last_op]
+
+    return mlp

pifuhd/lib/networks.py

@@ -0,0 +1,242 @@
+'''
+Copyright (C) 2019 NVIDIA Corporation. Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu.
+BSD License. All rights reserved. 
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+* Redistributions of source code must retain the above copyright notice, this
+  list of conditions and the following disclaimer.
+
+* Redistributions in binary form must reproduce the above copyright notice,
+  this list of conditions and the following disclaimer in the documentation
+  and/or other materials provided with the distribution.
+
+THE AUTHOR DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE, INCLUDING ALL 
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR ANY PARTICULAR PURPOSE. 
+IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY SPECIAL, INDIRECT OR CONSEQUENTIAL 
+DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, 
+WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING 
+OUT OF OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
+'''
+import torch
+import torch.nn as nn
+import functools
+from torch.autograd import Variable
+import numpy as np
+
+###############################################################################
+# Functions
+###############################################################################
+def weights_init(m):
+    classname = m.__class__.__name__
+    if classname.find('Conv') != -1:
+        m.weight.data.normal_(0.0, 0.02)
+    elif classname.find('BatchNorm2d') != -1:
+        m.weight.data.normal_(1.0, 0.02)
+        m.bias.data.fill_(0)
+
+def get_norm_layer(norm_type='instance'):
+    if norm_type == 'batch':
+        norm_layer = functools.partial(nn.BatchNorm2d, affine=True)
+    elif norm_type == 'instance':
+        norm_layer = functools.partial(nn.InstanceNorm2d, affine=False)
+    else:
+        raise NotImplementedError('normalization layer [%s] is not found' % norm_type)
+    return norm_layer
+
+def define_G(input_nc, output_nc, ngf, netG, n_downsample_global=3, n_blocks_global=9, n_local_enhancers=1, 
+             n_blocks_local=3, norm='instance', gpu_ids=[], last_op=nn.Tanh()):    
+    norm_layer = get_norm_layer(norm_type=norm)     
+    if netG == 'global':    
+        netG = GlobalGenerator(input_nc, output_nc, ngf, n_downsample_global, n_blocks_global, norm_layer, last_op=last_op)       
+    elif netG == 'local':        
+        netG = LocalEnhancer(input_nc, output_nc, ngf, n_downsample_global, n_blocks_global, 
+                                  n_local_enhancers, n_blocks_local, norm_layer)
+    elif netG == 'encoder':
+        netG = Encoder(input_nc, output_nc, ngf, n_downsample_global, norm_layer)
+    else:
+        raise('generator not implemented!')
+    # print(netG)
+    if len(gpu_ids) > 0:
+        assert(torch.cuda.is_available())   
+        netG.cuda(gpu_ids[0])
+    netG.apply(weights_init)
+    return netG
+
+def print_network(net):
+    if isinstance(net, list):
+        net = net[0]
+    num_params = 0
+    for param in net.parameters():
+        num_params += param.numel()
+    print(net)
+    print('Total number of parameters: %d' % num_params)
+
+##############################################################################
+# Generator
+##############################################################################
+class LocalEnhancer(nn.Module):
+    def __init__(self, input_nc, output_nc, ngf=32, n_downsample_global=3, n_blocks_global=9, 
+                 n_local_enhancers=1, n_blocks_local=3, norm_layer=nn.BatchNorm2d, padding_type='reflect'):        
+        super(LocalEnhancer, self).__init__()
+        self.n_local_enhancers = n_local_enhancers
+        
+        ###### global generator model #####           
+        ngf_global = ngf * (2**n_local_enhancers)
+        model_global = GlobalGenerator(input_nc, output_nc, ngf_global, n_downsample_global, n_blocks_global, norm_layer).model        
+        model_global = [model_global[i] for i in range(len(model_global)-3)] # get rid of final convolution layers        
+        self.model = nn.Sequential(*model_global)                
+
+        ###### local enhancer layers #####
+        for n in range(1, n_local_enhancers+1):
+            ### downsample            
+            ngf_global = ngf * (2**(n_local_enhancers-n))
+            model_downsample = [nn.ReflectionPad2d(3), nn.Conv2d(input_nc, ngf_global, kernel_size=7, padding=0), 
+                                norm_layer(ngf_global), nn.ReLU(True),
+                                nn.Conv2d(ngf_global, ngf_global * 2, kernel_size=3, stride=2, padding=1), 
+                                norm_layer(ngf_global * 2), nn.ReLU(True)]
+            ### residual blocks
+            model_upsample = []
+            for i in range(n_blocks_local):
+                model_upsample += [ResnetBlock(ngf_global * 2, padding_type=padding_type, norm_layer=norm_layer)]
+
+            ### upsample
+            model_upsample += [nn.ConvTranspose2d(ngf_global * 2, ngf_global, kernel_size=3, stride=2, padding=1, output_padding=1), 
+                               norm_layer(ngf_global), nn.ReLU(True)]      
+
+            ### final convolution
+            if n == n_local_enhancers:                
+                model_upsample += [nn.ReflectionPad2d(3), nn.Conv2d(ngf, output_nc, kernel_size=7, padding=0), nn.Tanh()]                       
+            
+            setattr(self, 'model'+str(n)+'_1', nn.Sequential(*model_downsample))
+            setattr(self, 'model'+str(n)+'_2', nn.Sequential(*model_upsample))                  
+        
+        self.downsample = nn.AvgPool2d(3, stride=2, padding=[1, 1], count_include_pad=False)
+
+    def forward(self, input): 
+        ### create input pyramid
+        input_downsampled = [input]
+        for i in range(self.n_local_enhancers):
+            input_downsampled.append(self.downsample(input_downsampled[-1]))
+
+        ### output at coarest level
+        output_prev = self.model(input_downsampled[-1])        
+        ### build up one layer at a time
+        for n_local_enhancers in range(1, self.n_local_enhancers+1):
+            model_downsample = getattr(self, 'model'+str(n_local_enhancers)+'_1')
+            model_upsample = getattr(self, 'model'+str(n_local_enhancers)+'_2')            
+            input_i = input_downsampled[self.n_local_enhancers-n_local_enhancers]            
+            output_prev = model_upsample(model_downsample(input_i) + output_prev)
+        return output_prev
+
+class GlobalGenerator(nn.Module):
+    def __init__(self, input_nc, output_nc, ngf=64, n_downsampling=3, n_blocks=9, norm_layer=nn.BatchNorm2d, 
+                 padding_type='reflect', last_op=nn.Tanh()):
+        assert(n_blocks >= 0)
+        super(GlobalGenerator, self).__init__()        
+        activation = nn.ReLU(True)        
+
+        model = [nn.ReflectionPad2d(3), nn.Conv2d(input_nc, ngf, kernel_size=7, padding=0), norm_layer(ngf), activation]
+        ### downsample
+        for i in range(n_downsampling):
+            mult = 2**i
+            model += [nn.Conv2d(ngf * mult, ngf * mult * 2, kernel_size=3, stride=2, padding=1),
+                      norm_layer(ngf * mult * 2), activation]
+
+        ### resnet blocks
+        mult = 2**n_downsampling
+        for i in range(n_blocks):
+            model += [ResnetBlock(ngf * mult, padding_type=padding_type, activation=activation, norm_layer=norm_layer)]
+        
+        ### upsample         
+        for i in range(n_downsampling):
+            mult = 2**(n_downsampling - i)
+            model += [nn.ConvTranspose2d(ngf * mult, int(ngf * mult / 2), kernel_size=3, stride=2, padding=1, output_padding=1),
+                       norm_layer(int(ngf * mult / 2)), activation]
+        model += [nn.ReflectionPad2d(3), nn.Conv2d(ngf, output_nc, kernel_size=7, padding=0)]
+        if last_op is not None:
+            model += [last_op]        
+        self.model = nn.Sequential(*model)
+            
+    def forward(self, input):
+        return self.model(input)             
+        
+# Define a resnet block
+class ResnetBlock(nn.Module):
+    def __init__(self, dim, padding_type, norm_layer, activation=nn.ReLU(True), use_dropout=False):
+        super(ResnetBlock, self).__init__()
+        self.conv_block = self.build_conv_block(dim, padding_type, norm_layer, activation, use_dropout)
+
+    def build_conv_block(self, dim, padding_type, norm_layer, activation, use_dropout):
+        conv_block = []
+        p = 0
+        if padding_type == 'reflect':
+            conv_block += [nn.ReflectionPad2d(1)]
+        elif padding_type == 'replicate':
+            conv_block += [nn.ReplicationPad2d(1)]
+        elif padding_type == 'zero':
+            p = 1
+        else:
+            raise NotImplementedError('padding [%s] is not implemented' % padding_type)
+
+        conv_block += [nn.Conv2d(dim, dim, kernel_size=3, padding=p),
+                       norm_layer(dim),
+                       activation]
+        if use_dropout:
+            conv_block += [nn.Dropout(0.5)]
+
+        p = 0
+        if padding_type == 'reflect':
+            conv_block += [nn.ReflectionPad2d(1)]
+        elif padding_type == 'replicate':
+            conv_block += [nn.ReplicationPad2d(1)]
+        elif padding_type == 'zero':
+            p = 1
+        else:
+            raise NotImplementedError('padding [%s] is not implemented' % padding_type)
+        conv_block += [nn.Conv2d(dim, dim, kernel_size=3, padding=p),
+                       norm_layer(dim)]
+
+        return nn.Sequential(*conv_block)
+
+    def forward(self, x):
+        out = x + self.conv_block(x)
+        return out
+
+class Encoder(nn.Module):
+    def __init__(self, input_nc, output_nc, ngf=32, n_downsampling=4, norm_layer=nn.BatchNorm2d):
+        super(Encoder, self).__init__()        
+        self.output_nc = output_nc        
+
+        model = [nn.ReflectionPad2d(3), nn.Conv2d(input_nc, ngf, kernel_size=7, padding=0), 
+                 norm_layer(ngf), nn.ReLU(True)]             
+        ### downsample
+        for i in range(n_downsampling):
+            mult = 2**i
+            model += [nn.Conv2d(ngf * mult, ngf * mult * 2, kernel_size=3, stride=2, padding=1),
+                      norm_layer(ngf * mult * 2), nn.ReLU(True)]
+
+        ### upsample         
+        for i in range(n_downsampling):
+            mult = 2**(n_downsampling - i)
+            model += [nn.ConvTranspose2d(ngf * mult, int(ngf * mult / 2), kernel_size=3, stride=2, padding=1, output_padding=1),
+                       norm_layer(int(ngf * mult / 2)), nn.ReLU(True)]        
+
+        model += [nn.ReflectionPad2d(3), nn.Conv2d(ngf, output_nc, kernel_size=7, padding=0), nn.Tanh()]
+        self.model = nn.Sequential(*model) 
+
+    def forward(self, input, inst):
+        outputs = self.model(input)
+
+        # instance-wise average pooling
+        outputs_mean = outputs.clone()
+        inst_list = np.unique(inst.cpu().numpy().astype(int))        
+        for i in inst_list:
+            for b in range(input.size()[0]):
+                indices = (inst[b:b+1] == int(i)).nonzero() # n x 4            
+                for j in range(self.output_nc):
+                    output_ins = outputs[indices[:,0] + b, indices[:,1] + j, indices[:,2], indices[:,3]]                    
+                    mean_feat = torch.mean(output_ins).expand_as(output_ins)                                        
+                    outputs_mean[indices[:,0] + b, indices[:,1] + j, indices[:,2], indices[:,3]] = mean_feat                       
+        return outputs_mean

pifuhd/lib/options.py

@@ -0,0 +1,208 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All rights reserved.
+
+import argparse
+import os
+
+class BaseOptions():
+    def __init__(self):
+        self.initialized = False
+        self.parser = None
+
+    def initialize(self, parser):
+        # Datasets related
+        g_data = parser.add_argument_group('Data')
+        g_data.add_argument('--dataset', type=str, default='renderppl', help='dataset name')
+        g_data.add_argument('--dataroot', type=str, default='./data',
+                            help='path to images (data folder)')
+
+        g_data.add_argument('--loadSize', type=int, default=512, help='load size of input image')
+
+        # Experiment related
+        g_exp = parser.add_argument_group('Experiment')
+        g_exp.add_argument('--name', type=str, default='',
+                           help='name of the experiment. It decides where to store samples and models')
+        g_exp.add_argument('--debug', action='store_true', help='debug mode or not')
+        g_exp.add_argument('--mode', type=str, default='inout', help='inout || color')
+
+        # Training related
+        g_train = parser.add_argument_group('Training')
+        g_train.add_argument('--tmp_id', type=int, default=0, help='tmp_id')
+        g_train.add_argument('--gpu_id', type=int, default=0, help='gpu id for cuda')
+        g_train.add_argument('--batch_size', type=int, default=32, help='input batch size')
+        g_train.add_argument('--num_threads', default=1, type=int, help='# sthreads for loading data')
+        g_train.add_argument('--serial_batches', action='store_true',
+                             help='if true, takes images in order to make batches, otherwise takes them randomly')
+        g_train.add_argument('--pin_memory', action='store_true', help='pin_memory')
+        g_train.add_argument('--learning_rate', type=float, default=1e-3, help='adam learning rate')
+        g_train.add_argument('--num_iter', type=int, default=30000, help='num iterations to train')
+        g_train.add_argument('--freq_plot', type=int, default=100, help='freqency of the error plot')
+        g_train.add_argument('--freq_mesh', type=int, default=20000, help='freqency of the save_checkpoints')
+        g_train.add_argument('--freq_eval', type=int, default=5000, help='freqency of the save_checkpoints')
+        g_train.add_argument('--freq_save_ply', type=int, default=5000, help='freqency of the save ply')
+        g_train.add_argument('--freq_save_image', type=int, default=100, help='freqency of the save input image')
+        g_train.add_argument('--resume_epoch', type=int, default=-1, help='epoch resuming the training')
+        g_train.add_argument('--continue_train', action='store_true', help='continue training: load the latest model')
+        g_train.add_argument('--finetune', action='store_true', help='fine tuning netG in training C')
+
+        # Testing related
+        g_test = parser.add_argument_group('Testing')
+        g_test.add_argument('--resolution', type=int, default=512, help='# of grid in mesh reconstruction')
+        g_test.add_argument('--no_numel_eval', action='store_true', help='no numerical evaluation')
+        g_test.add_argument('--no_mesh_recon', action='store_true', help='no mesh reconstruction')
+
+        # Sampling related
+        g_sample = parser.add_argument_group('Sampling')
+        g_sample.add_argument('--num_sample_inout', type=int, default=6000, help='# of sampling points')
+        g_sample.add_argument('--num_sample_surface', type=int, default=0, help='# of sampling points')
+        g_sample.add_argument('--num_sample_normal', type=int, default=0, help='# of sampling points')
+        g_sample.add_argument('--num_sample_color', type=int, default=0, help='# of sampling points')
+        g_sample.add_argument('--num_pts_dic', type=int, default=1, help='# of pts dic you load')
+
+        g_sample.add_argument('--crop_type', type=str, default='fullbody', help='Sampling file name.')
+        g_sample.add_argument('--uniform_ratio', type=float, default=0.1, help='maximum sigma for sampling')
+        g_sample.add_argument('--mask_ratio', type=float, default=0.5, help='maximum sigma for sampling')
+        g_sample.add_argument('--sampling_parts', action='store_true', help='Sampling on the fly')
+        g_sample.add_argument('--sampling_otf', action='store_true', help='Sampling on the fly')
+        g_sample.add_argument('--sampling_mode', type=str, default='sigma_uniform', help='Sampling file name.')
+        g_sample.add_argument('--linear_anneal_sigma', action='store_true', help='linear annealing of sigma')
+        g_sample.add_argument('--sigma_max', type=float, default=0.0, help='maximum sigma for sampling')
+        g_sample.add_argument('--sigma_min', type=float, default=0.0, help='minimum sigma for sampling')
+        g_sample.add_argument('--sigma', type=float, default=1.0, help='sigma for sampling')
+        g_sample.add_argument('--sigma_surface', type=float, default=1.0, help='sigma for sampling')
+        
+        g_sample.add_argument('--z_size', type=float, default=200.0, help='z normalization factor')
+
+        # Model related
+        g_model = parser.add_argument_group('Model')
+        # General
+        g_model.add_argument('--norm', type=str, default='batch',
+                             help='instance normalization or batch normalization or group normalization')
+
+        # Image filter General
+        g_model.add_argument('--netG', type=str, default='hgpifu', help='piximp | fanimp | hghpifu')
+        g_model.add_argument('--netC', type=str, default='resblkpifu', help='resblkpifu | resblkhpifu')
+
+        # hgimp specific
+        g_model.add_argument('--num_stack', type=int, default=4, help='# of hourglass')
+        g_model.add_argument('--hg_depth', type=int, default=2, help='# of stacked layer of hourglass')
+        g_model.add_argument('--hg_down', type=str, default='ave_pool', help='ave pool || conv64 || conv128')
+        g_model.add_argument('--hg_dim', type=int, default=256, help='256 | 512')
+
+        # Classification General
+        g_model.add_argument('--mlp_norm', type=str, default='group', help='normalization for volume branch')
+        g_model.add_argument('--mlp_dim', nargs='+', default=[257, 1024, 512, 256, 128, 1], type=int,
+                             help='# of dimensions of mlp. no need to put the first channel')
+        g_model.add_argument('--mlp_dim_color', nargs='+', default=[1024, 512, 256, 128, 3], type=int,
+                             help='# of dimensions of mlp. no need to put the first channel')
+        g_model.add_argument('--mlp_res_layers', nargs='+', default=[2,3,4], type=int,
+                             help='leyers that has skip connection. use 0 for no residual pass')
+        g_model.add_argument('--merge_layer', type=int, default=-1)
+
+        # for train
+        parser.add_argument('--random_body_chop', action='store_true', help='if random flip')
+        parser.add_argument('--random_flip', action='store_true', help='if random flip')
+        parser.add_argument('--random_trans', action='store_true', help='if random flip')
+        parser.add_argument('--random_scale', action='store_true', help='if random flip')
+        parser.add_argument('--random_rotate', action='store_true', help='if random flip')
+        parser.add_argument('--random_bg', action='store_true', help='using random background')
+
+        parser.add_argument('--schedule', type=int, nargs='+', default=[10, 15],
+                            help='Decrease learning rate at these epochs.')
+        parser.add_argument('--gamma', type=float, default=0.1, help='LR is multiplied by gamma on schedule.')
+        parser.add_argument('--lambda_nml', type=float, default=0.0, help='weight of normal loss')
+        parser.add_argument('--lambda_cmp_l1', type=float, default=0.0, help='weight of normal loss')
+        parser.add_argument('--occ_loss_type', type=str, default='mse', help='bce | brock_bce | mse')
+        parser.add_argument('--clr_loss_type', type=str, default='mse', help='mse | l1')
+        parser.add_argument('--nml_loss_type', type=str, default='mse', help='mse | l1')
+        parser.add_argument('--occ_gamma', type=float, default=None, help='weighting term')
+        parser.add_argument('--no_finetune', action='store_true', help='fine tuning netG in training C')
+
+        # for eval
+        parser.add_argument('--val_test_error', action='store_true', help='validate errors of test data')
+        parser.add_argument('--val_train_error', action='store_true', help='validate errors of train data')
+        parser.add_argument('--gen_test_mesh', action='store_true', help='generate test mesh')
+        parser.add_argument('--gen_train_mesh', action='store_true', help='generate train mesh')
+        parser.add_argument('--all_mesh', action='store_true', help='generate meshs from all hourglass output')
+        parser.add_argument('--num_gen_mesh_test', type=int, default=4,
+                            help='how many meshes to generate during testing')
+
+        # path
+        parser.add_argument('--load_netG_checkpoint_path', type=str, help='path to save checkpoints')
+        parser.add_argument('--load_netC_checkpoint_path', type=str, help='path to save checkpoints')
+        parser.add_argument('--checkpoints_path', type=str, default='./checkpoints', help='path to save checkpoints')
+        parser.add_argument('--results_path', type=str, default='./results', help='path to save results ply')
+        parser.add_argument('--load_checkpoint_path', type=str, help='path to save results ply')
+        parser.add_argument('--single', type=str, default='', help='single data for training')
+        
+        # for single image reconstruction
+        parser.add_argument('--mask_path', type=str, help='path for input mask')
+        parser.add_argument('--img_path', type=str, help='path for input image')
+
+        # for multi resolution
+        parser.add_argument('--load_netMR_checkpoint_path', type=str, help='path to save checkpoints')
+        parser.add_argument('--loadSizeBig', type=int, default=1024, help='load size of input image')
+        parser.add_argument('--loadSizeLocal', type=int, default=512, help='load size of input image')
+        parser.add_argument('--train_full_pifu', action='store_true', help='enable end-to-end training')
+        parser.add_argument('--num_local', type=int, default=1, help='number of local cropping')
+
+        # for normal condition
+        parser.add_argument('--load_netFB_checkpoint_path', type=str, help='path to save checkpoints')
+        parser.add_argument('--load_netF_checkpoint_path', type=str, help='path to save checkpoints')
+        parser.add_argument('--load_netB_checkpoint_path', type=str, help='path to save checkpoints')
+        parser.add_argument('--use_aio_normal', action='store_true')
+        parser.add_argument('--use_front_normal', action='store_true')
+        parser.add_argument('--use_back_normal', action='store_true')
+        parser.add_argument('--no_intermediate_loss', action='store_true')
+
+        # aug
+        group_aug = parser.add_argument_group('aug')
+        group_aug.add_argument('--aug_alstd', type=float, default=0.0, help='augmentation pca lighting alpha std')
+        group_aug.add_argument('--aug_bri', type=float, default=0.2, help='augmentation brightness')
+        group_aug.add_argument('--aug_con', type=float, default=0.2, help='augmentation contrast')
+        group_aug.add_argument('--aug_sat', type=float, default=0.05, help='augmentation saturation')
+        group_aug.add_argument('--aug_hue', type=float, default=0.05, help='augmentation hue')
+        group_aug.add_argument('--aug_gry', type=float, default=0.1, help='augmentation gray scale')
+        group_aug.add_argument('--aug_blur', type=float, default=0.0, help='augmentation blur')
+
+        # for reconstruction
+        parser.add_argument('--start_id', type=int, default=-1, help='load size of input image')
+        parser.add_argument('--end_id', type=int, default=-1, help='load size of input image')
+
+        # special tasks
+        self.initialized = True
+        return parser
+
+    def gather_options(self, args=None):
+        # initialize parser with basic options
+        if not self.initialized:
+            parser = argparse.ArgumentParser(
+                formatter_class=argparse.ArgumentDefaultsHelpFormatter)
+            parser = self.initialize(parser)
+            self.parser = parser
+
+        if args is None:
+            return self.parser.parse_args()
+        else:
+            return self.parser.parse_args(args)
+
+    def print_options(self, opt):
+        message = ''
+        message += '----------------- Options ---------------\n'
+        for k, v in sorted(vars(opt).items()):
+            comment = ''
+            default = self.parser.get_default(k)
+            if v != default:
+                comment = '\t[default: %s]' % str(default)
+            message += '{:>25}: {:<30}{}\n'.format(str(k), str(v), comment)
+        message += '----------------- End -------------------'
+        print(message)
+
+    def parse(self, args=None):
+        opt = self.gather_options(args)
+
+        opt.sigma = opt.sigma_max
+
+        if len(opt.mlp_res_layers) == 1 and opt.mlp_res_layers[0] < 1:
+            opt.mlp_res_layers = []
+                
+        return opt

pifuhd/lib/render/__init__.py

@@ -0,0 +1 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All rights reserved.

pifuhd/lib/render/camera.py

@@ -0,0 +1,240 @@
+'''
+MIT License
+
+Copyright (c) 2019 Shunsuke Saito, Zeng Huang, and Ryota Natsume
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+'''
+import cv2
+import numpy as np
+
+from .glm import ortho
+
+
+class Camera:
+    def __init__(self, width=1600, height=1200):
+        # Focal Length
+        # equivalent 50mm
+        focal = np.sqrt(width * width + height * height)
+        self.focal_x = focal
+        self.focal_y = focal
+        # Principal Point Offset
+        self.principal_x = width / 2
+        self.principal_y = height / 2
+        # Axis Skew
+        self.skew = 0
+        # Image Size
+        self.width = width
+        self.height = height
+
+        self.near = 1
+        self.far = 10
+
+        # Camera Center
+        self.eye = np.array([0, 0, -3.6])
+        self.center = np.array([0, 0, 0])
+        self.direction = np.array([0, 0, -1])
+        self.right = np.array([1, 0, 0])
+        self.up = np.array([0, 1, 0])
+
+        self.ortho_ratio = None
+
+    def sanity_check(self):
+        self.center = self.center.reshape([-1])
+        self.direction = self.direction.reshape([-1])
+        self.right = self.right.reshape([-1])
+        self.up = self.up.reshape([-1])
+
+        assert len(self.center) == 3
+        assert len(self.direction) == 3
+        assert len(self.right) == 3
+        assert len(self.up) == 3
+
+    @staticmethod
+    def normalize_vector(v):
+        v_norm = np.linalg.norm(v)
+        return v if v_norm == 0 else v / v_norm
+
+    def get_real_z_value(self, z):
+        z_near = self.near
+        z_far = self.far
+        z_n = 2.0 * z - 1.0
+        z_e = 2.0 * z_near * z_far / (z_far + z_near - z_n * (z_far - z_near))
+        return z_e
+
+    def get_rotation_matrix(self):
+        rot_mat = np.eye(3)
+        d = self.eye - self.center
+        d = -self.normalize_vector(d)
+        u = self.up
+        self.right = -np.cross(u, d)
+        u = np.cross(d, self.right)
+        rot_mat[0, :] = self.right
+        rot_mat[1, :] = u
+        rot_mat[2, :] = d
+
+        # s = self.right
+        # s = self.normalize_vector(s)
+        # rot_mat[0, :] = s
+        # u = self.up
+        # u = self.normalize_vector(u)
+        # rot_mat[1, :] = -u
+        # rot_mat[2, :] = self.normalize_vector(self.direction)
+
+        return rot_mat
+
+    def get_translation_vector(self):
+        rot_mat = self.get_rotation_matrix()
+        trans = -np.dot(rot_mat.T, self.eye)
+        return trans
+
+    def get_intrinsic_matrix(self):
+        int_mat = np.eye(3)
+
+        int_mat[0, 0] = self.focal_x
+        int_mat[1, 1] = self.focal_y
+        int_mat[0, 1] = self.skew
+        int_mat[0, 2] = self.principal_x
+        int_mat[1, 2] = self.principal_y
+
+        return int_mat
+
+    def get_projection_matrix(self):
+        ext_mat = self.get_extrinsic_matrix()
+        int_mat = self.get_intrinsic_matrix()
+
+        return np.matmul(int_mat, ext_mat)
+
+    def get_extrinsic_matrix(self):
+        rot_mat = self.get_rotation_matrix()
+        int_mat = self.get_intrinsic_matrix()
+        trans = self.get_translation_vector()
+
+        extrinsic = np.eye(4)
+        extrinsic[:3, :3] = rot_mat
+        extrinsic[:3, 3] = trans
+
+        return extrinsic[:3, :]
+
+    def set_rotation_matrix(self, rot_mat):
+        self.direction = rot_mat[2, :]
+        self.up = -rot_mat[1, :]
+        self.right = rot_mat[0, :]
+
+    def set_intrinsic_matrix(self, int_mat):
+        self.focal_x = int_mat[0, 0]
+        self.focal_y = int_mat[1, 1]
+        self.skew = int_mat[0, 1]
+        self.principal_x = int_mat[0, 2]
+        self.principal_y = int_mat[1, 2]
+
+    def set_projection_matrix(self, proj_mat):
+        res = cv2.decomposeProjectionMatrix(proj_mat)
+        int_mat, rot_mat, camera_center_homo = res[0], res[1], res[2]
+        camera_center = camera_center_homo[0:3] / camera_center_homo[3]
+        camera_center = camera_center.reshape(-1)
+        int_mat = int_mat / int_mat[2][2]
+
+        self.set_intrinsic_matrix(int_mat)
+        self.set_rotation_matrix(rot_mat)
+        self.center = camera_center
+
+        self.sanity_check()
+
+    def get_gl_matrix(self):
+        z_near = self.near
+        z_far = self.far
+        rot_mat = self.get_rotation_matrix()
+        int_mat = self.get_intrinsic_matrix()
+        trans = self.get_translation_vector()
+
+        extrinsic = np.eye(4)
+        extrinsic[:3, :3] = rot_mat
+        extrinsic[:3, 3] = trans
+        axis_adj = np.eye(4)
+        axis_adj[2, 2] = -1
+        axis_adj[1, 1] = -1
+        model_view = np.matmul(axis_adj, extrinsic)
+
+        projective = np.zeros([4, 4])
+        projective[:2, :2] = int_mat[:2, :2]
+        projective[:2, 2:3] = -int_mat[:2, 2:3]
+        projective[3, 2] = -1
+        projective[2, 2] = (z_near + z_far)
+        projective[2, 3] = (z_near * z_far)
+
+        if self.ortho_ratio is None:
+            ndc = ortho(0, self.width, 0, self.height, z_near, z_far)
+            perspective = np.matmul(ndc, projective)
+        else:
+            perspective = ortho(-self.width * self.ortho_ratio / 2, self.width * self.ortho_ratio / 2,
+                                -self.height * self.ortho_ratio / 2, self.height * self.ortho_ratio / 2,
+                                z_near, z_far)
+
+        return perspective, model_view
+
+
+def KRT_from_P(proj_mat, normalize_K=True):
+    res = cv2.decomposeProjectionMatrix(proj_mat)
+    K, Rot, camera_center_homog = res[0], res[1], res[2]
+    camera_center = camera_center_homog[0:3] / camera_center_homog[3]
+    trans = -Rot.dot(camera_center)
+    if normalize_K:
+        K = K / K[2][2]
+    return K, Rot, trans
+
+
+def MVP_from_P(proj_mat, width, height, near=0.1, far=10000):
+    '''
+    Convert OpenCV camera calibration matrix to OpenGL projection and model view matrix
+    :param proj_mat: OpenCV camera projeciton matrix
+    :param width: Image width
+    :param height: Image height
+    :param near: Z near value
+    :param far: Z far value
+    :return: OpenGL projection matrix and model view matrix
+    '''
+    res = cv2.decomposeProjectionMatrix(proj_mat)
+    K, Rot, camera_center_homog = res[0], res[1], res[2]
+    camera_center = camera_center_homog[0:3] / camera_center_homog[3]
+    trans = -Rot.dot(camera_center)
+    K = K / K[2][2]
+
+    extrinsic = np.eye(4)
+    extrinsic[:3, :3] = Rot
+    extrinsic[:3, 3:4] = trans
+    axis_adj = np.eye(4)
+    axis_adj[2, 2] = -1
+    axis_adj[1, 1] = -1
+    model_view = np.matmul(axis_adj, extrinsic)
+
+    zFar = far
+    zNear = near
+    projective = np.zeros([4, 4])
+    projective[:2, :2] = K[:2, :2]
+    projective[:2, 2:3] = -K[:2, 2:3]
+    projective[3, 2] = -1
+    projective[2, 2] = (zNear + zFar)
+    projective[2, 3] = (zNear * zFar)
+
+    ndc = ortho(0, width, 0, height, zNear, zFar)
+
+    perspective = np.matmul(ndc, projective)
+
+    return perspective, model_view

pifuhd/lib/render/gl/__init__.py

@@ -0,0 +1,26 @@
+'''
+MIT License
+
+Copyright (c) 2019 Shunsuke Saito, Zeng Huang, and Ryota Natsume
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+'''
+from .framework import *
+from .render import *
+from .cam_render import *

pifuhd/lib/render/gl/cam_render.py

@@ -0,0 +1,72 @@
+'''
+MIT License
+
+Copyright (c) 2019 Shunsuke Saito, Zeng Huang, and Ryota Natsume
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+'''
+from OpenGL.GLUT import *
+
+from .render import Render
+
+
+class CamRender(Render):
+    def __init__(self, width=1600, height=1200, name='Cam Renderer',
+                 program_files=['simple.fs', 'simple.vs'], color_size=1, ms_rate=1):
+        Render.__init__(self, width, height, name, program_files, color_size, ms_rate)
+        self.camera = None
+
+        glutDisplayFunc(self.display)
+        glutKeyboardFunc(self.keyboard)
+
+    def set_camera(self, camera):
+        self.camera = camera
+        self.projection_matrix, self.model_view_matrix = camera.get_gl_matrix()
+
+    def set_matrices(self, projection, modelview):
+        self.projection_matrix = projection
+        self.model_view_matrix = modelview
+
+    def keyboard(self, key, x, y):
+        # up
+        eps = 1
+        # print(key)
+        if key == b'w':
+            self.camera.center += eps * self.camera.direction
+        elif key == b's':
+            self.camera.center -= eps * self.camera.direction
+        if key == b'a':
+            self.camera.center -= eps * self.camera.right
+        elif key == b'd':
+            self.camera.center += eps * self.camera.right
+        if key == b' ':
+            self.camera.center += eps * self.camera.up
+        elif key == b'x':
+            self.camera.center -= eps * self.camera.up
+        elif key == b'i':
+            self.camera.near += 0.1 * eps
+            self.camera.far += 0.1 * eps
+        elif key == b'o':
+            self.camera.near -= 0.1 * eps
+            self.camera.far -= 0.1 * eps
+
+        self.projection_matrix, self.model_view_matrix = self.camera.get_gl_matrix()
+
+    def show(self):
+        glutMainLoop()

pifuhd/lib/render/gl/color_render.py

@@ -0,0 +1,113 @@
+'''
+MIT License
+
+Copyright (c) 2019 Shunsuke Saito, Zeng Huang, and Ryota Natsume
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+'''
+import numpy as np
+import random
+
+from .framework import *
+from .cam_render import CamRender
+
+
+class ColorRender(CamRender):
+    def __init__(self, width=1600, height=1200, name='Color Renderer'):
+        program_files = ['color.vs', 'color.fs']
+        CamRender.__init__(self, width, height, name, program_files=program_files)
+
+        # WARNING: this differs from vertex_buffer and vertex_data in Render
+        self.vert_buffer = {}
+        self.vert_data = {}
+
+        self.color_buffer = {}
+        self.color_data = {}
+
+        self.vertex_dim = {}
+        self.n_vertices = {}
+
+    def set_mesh(self, vertices, faces, color, faces_clr, mat_name='all'):
+        self.vert_data[mat_name] = vertices[faces.reshape([-1])]
+        self.n_vertices[mat_name] = self.vert_data[mat_name].shape[0]
+        self.vertex_dim[mat_name] = self.vert_data[mat_name].shape[1]
+
+        if mat_name not in self.vert_buffer.keys():
+            self.vert_buffer[mat_name] = glGenBuffers(1)
+        glBindBuffer(GL_ARRAY_BUFFER, self.vert_buffer[mat_name])
+        glBufferData(GL_ARRAY_BUFFER, self.vert_data[mat_name], GL_STATIC_DRAW)
+
+        self.color_data[mat_name] = color[faces_clr.reshape([-1])]
+        if mat_name not in self.color_buffer.keys():
+            self.color_buffer[mat_name] = glGenBuffers(1)
+        glBindBuffer(GL_ARRAY_BUFFER, self.color_buffer[mat_name])
+        glBufferData(GL_ARRAY_BUFFER, self.color_data[mat_name], GL_STATIC_DRAW)
+
+        glBindBuffer(GL_ARRAY_BUFFER, 0)
+
+    def cleanup(self):
+        
+        glBindBuffer(GL_ARRAY_BUFFER, 0)
+        for key in self.vert_data:
+            glDeleteBuffers(1, [self.vert_buffer[key]])
+            glDeleteBuffers(1, [self.color_buffer[key]])
+
+        self.vert_buffer = {}
+        self.vert_data = {}
+    
+        self.color_buffer = {}
+        self.color_data = {}
+
+        self.render_texture_mat = {}
+
+        self.vertex_dim = {}
+        self.n_vertices = {}
+
+    def draw(self):
+        self.draw_init()
+
+        glEnable(GL_MULTISAMPLE)
+
+        glUseProgram(self.program)
+        glUniformMatrix4fv(self.model_mat_unif, 1, GL_FALSE, self.model_view_matrix.transpose())
+        glUniformMatrix4fv(self.persp_mat_unif, 1, GL_FALSE, self.projection_matrix.transpose())
+
+        for mat in self.vert_buffer:
+            # Handle vertex buffer
+            glBindBuffer(GL_ARRAY_BUFFER, self.vert_buffer[mat])
+            glEnableVertexAttribArray(0)
+            glVertexAttribPointer(0, self.vertex_dim[mat], GL_DOUBLE, GL_FALSE, 0, None)
+
+            # Handle normal buffer
+            glBindBuffer(GL_ARRAY_BUFFER, self.color_buffer[mat])
+            glEnableVertexAttribArray(1)
+            glVertexAttribPointer(1, 3, GL_DOUBLE, GL_FALSE, 0, None)
+
+            glDrawArrays(GL_TRIANGLES, 0, self.n_vertices[mat])
+
+            glDisableVertexAttribArray(1)
+            glDisableVertexAttribArray(0)
+
+        glBindBuffer(GL_ARRAY_BUFFER, 0)
+
+        glUseProgram(0)
+
+        glDisable(GL_MULTISAMPLE)
+
+        self.draw_end()

pifuhd/lib/render/gl/data/color.fs

@@ -0,0 +1,10 @@
+#version 330 core
+
+out vec4 FragColor;
+
+in vec3 Color;
+
+void main() 
+{
+    FragColor = vec4(Color,1.0);
+}

pifuhd/lib/render/gl/data/color.vs

@@ -0,0 +1,17 @@
+#version 330 core
+
+layout (location = 0) in vec3 a_Position;
+layout (location = 1) in vec3 a_Color;
+
+out vec3 CamNormal;
+out vec3 CamPos;
+out vec3 Color;
+
+uniform mat4 ModelMat;
+uniform mat4 PerspMat;
+
+void main()
+{
+	gl_Position = PerspMat * ModelMat * vec4(a_Position, 1.0);
+    Color = a_Color;
+}

pifuhd/lib/render/gl/data/geo.fs

@@ -0,0 +1,14 @@
+#version 330 core
+
+out vec4 FragColor;
+
+in vec3 CamNormal;
+in vec3 CamPos;
+
+void main() 
+{
+    vec3 light_direction = vec3(0, 0, 1);
+    vec3 f_normal = normalize(CamNormal.xyz);
+    vec4 specular_reflection = vec4(0.2) * pow(max(0.0, dot(reflect(-light_direction, f_normal), vec3(0, 0, -1))), 16.f);
+    FragColor = vec4(dot(f_normal, light_direction)*vec3(1.0)+specular_reflection.xyz, 1.0);
+}

pifuhd/lib/render/gl/data/geo.vs

@@ -0,0 +1,17 @@
+#version 330 core
+
+layout (location = 0) in vec3 a_Position;
+layout (location = 1) in vec3 a_Normal;
+
+out vec3 CamNormal;
+out vec3 CamPos;
+
+uniform mat4 ModelMat;
+uniform mat4 PerspMat;
+
+void main()
+{
+	gl_Position = PerspMat * ModelMat * vec4(a_Position, 1.0);
+    CamNormal = (ModelMat * vec4(a_Normal, 0.0)).xyz;
+    CamPos = (ModelMat * vec4(a_Position, 1.0)).xyz;
+}

pifuhd/lib/render/gl/data/normal.fs

@@ -0,0 +1,12 @@
+#version 330
+
+out vec4 FragColor;
+
+in vec3 CamNormal;
+
+void main()
+{
+    vec3 cam_norm_normalized = normalize(CamNormal);
+    vec3 rgb = (cam_norm_normalized + 1.0) / 2.0;
+	FragColor = vec4(rgb, 1.0);
+}

pifuhd/lib/render/gl/data/normal.vs

@@ -0,0 +1,15 @@
+#version 330
+
+layout (location = 0) in vec3 Position;
+layout (location = 1) in vec3 Normal;
+
+out vec3 CamNormal;
+
+uniform mat4 ModelMat;
+uniform mat4 PerspMat;
+
+void main()
+{
+	gl_Position = PerspMat * ModelMat * vec4(Position, 1.0);
+	CamNormal = (ModelMat * vec4(Normal, 0.0)).xyz;
+}

pifuhd/lib/render/gl/data/quad.fs

@@ -0,0 +1,12 @@
+#version 330 core
+
+out vec4 FragColor;
+
+in vec2 TexCoord;
+
+uniform sampler2D screenTexture;
+
+void main()
+{
+    FragColor = texture(screenTexture, TexCoord);
+}

pifuhd/lib/render/gl/data/quad.vs

@@ -0,0 +1,12 @@
+#version 330 core
+
+layout (location = 0) in vec2 aPos;
+layout (location = 1) in vec2 aTexCoord;
+
+out vec2 TexCoord;
+
+void main()
+{
+    gl_Position = vec4(aPos.x, aPos.y, 0.0, 1.0);
+    TexCoord = aTexCoord;
+}

pifuhd/lib/render/gl/framework.py

@@ -0,0 +1,93 @@
+# Mario Rosasco, 2016
+# adapted from framework.cpp, Copyright (C) 2010-2012 by Jason L. McKesson
+# This file is licensed under the MIT License.
+#
+# NB: Unlike in the framework.cpp organization, the main loop is contained
+# in the tutorial files, not in this framework file. Additionally, a copy of
+# this module file must exist in the same directory as the tutorial files
+# to be imported properly.
+
+
+import os
+
+from OpenGL.GL import *
+
+
+# Function that creates and compiles shaders according to the given type (a GL enum value) and
+# shader program (a file containing a GLSL program).
+def loadShader(shaderType, shaderFile):
+    # check if file exists, get full path name
+    strFilename = findFileOrThrow(shaderFile)
+    shaderData = None
+    with open(strFilename, 'r') as f:
+        shaderData = f.read()
+
+    shader = glCreateShader(shaderType)
+    glShaderSource(shader, shaderData)  # note that this is a simpler function call than in C
+
+    # This shader compilation is more explicit than the one used in
+    # framework.cpp, which relies on a glutil wrapper function.
+    # This is made explicit here mainly to decrease dependence on pyOpenGL
+    # utilities and wrappers, which docs caution may change in future versions.
+    glCompileShader(shader)
+
+    status = glGetShaderiv(shader, GL_COMPILE_STATUS)
+    if status == GL_FALSE:
+        # Note that getting the error log is much simpler in Python than in C/C++
+        # and does not require explicit handling of the string buffer
+        strInfoLog = glGetShaderInfoLog(shader)
+        strShaderType = ""
+        if shaderType is GL_VERTEX_SHADER:
+            strShaderType = "vertex"
+        elif shaderType is GL_GEOMETRY_SHADER:
+            strShaderType = "geometry"
+        elif shaderType is GL_FRAGMENT_SHADER:
+            strShaderType = "fragment"
+
+        print("Compilation failure for " + strShaderType + " shader:\n" + str(strInfoLog))
+
+    return shader
+
+
+# Function that accepts a list of shaders, compiles them, and returns a handle to the compiled program
+def createProgram(shaderList):
+    program = glCreateProgram()
+
+    for shader in shaderList:
+        glAttachShader(program, shader)
+
+    glLinkProgram(program)
+
+    status = glGetProgramiv(program, GL_LINK_STATUS)
+    if status == GL_FALSE:
+        # Note that getting the error log is much simpler in Python than in C/C++
+        # and does not require explicit handling of the string buffer
+        strInfoLog = glGetProgramInfoLog(program)
+        print("Linker failure: \n" + str(strInfoLog))
+
+    for shader in shaderList:
+        glDetachShader(program, shader)
+
+    return program
+
+
+# Helper function to locate and open the target file (passed in as a string).
+# Returns the full path to the file as a string.
+def findFileOrThrow(strBasename):
+    # Keep constant names in C-style convention, for readability
+    # when comparing to C(/C++) code.
+    if os.path.isfile(strBasename):
+        return strBasename
+
+    LOCAL_FILE_DIR = "data" + os.sep
+    GLOBAL_FILE_DIR = os.path.dirname(os.path.abspath(__file__)) + os.sep + "data" + os.sep
+
+    strFilename = LOCAL_FILE_DIR + strBasename
+    if os.path.isfile(strFilename):
+        return strFilename
+
+    strFilename = GLOBAL_FILE_DIR + strBasename
+    if os.path.isfile(strFilename):
+        return strFilename
+
+    raise IOError('Could not find target file ' + strBasename)

pifuhd/lib/render/gl/geo_render.py

@@ -0,0 +1,113 @@
+'''
+MIT License
+
+Copyright (c) 2019 Shunsuke Saito, Zeng Huang, and Ryota Natsume
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+'''
+import numpy as np
+import random
+
+from .framework import *
+from .cam_render import CamRender
+
+
+class GeoRender(CamRender):
+    def __init__(self, width=1600, height=1200, name='Geo Renderer'):
+        program_files = ['geo.vs', 'geo.fs']
+        CamRender.__init__(self, width, height, name, program_files=program_files)
+
+        # WARNING: this differs from vertex_buffer and vertex_data in Render
+        self.vert_buffer = {}
+        self.vert_data = {}
+
+        self.norm_buffer = {}
+        self.norm_data = {}
+
+        self.vertex_dim = {}
+        self.n_vertices = {}
+
+    def set_mesh(self, vertices, faces, norms, faces_nml, mat_name='all'):
+        self.vert_data[mat_name] = vertices[faces.reshape([-1])]
+        self.n_vertices[mat_name] = self.vert_data[mat_name].shape[0]
+        self.vertex_dim[mat_name] = self.vert_data[mat_name].shape[1]
+
+        if mat_name not in self.vert_buffer.keys():
+            self.vert_buffer[mat_name] = glGenBuffers(1)
+        glBindBuffer(GL_ARRAY_BUFFER, self.vert_buffer[mat_name])
+        glBufferData(GL_ARRAY_BUFFER, self.vert_data[mat_name], GL_STATIC_DRAW)
+
+        self.norm_data[mat_name] = norms[faces_nml.reshape([-1])]
+        if mat_name not in self.norm_buffer.keys():
+            self.norm_buffer[mat_name] = glGenBuffers(1)
+        glBindBuffer(GL_ARRAY_BUFFER, self.norm_buffer[mat_name])
+        glBufferData(GL_ARRAY_BUFFER, self.norm_data[mat_name], GL_STATIC_DRAW)
+
+        glBindBuffer(GL_ARRAY_BUFFER, 0)
+
+    def cleanup(self):
+        
+        glBindBuffer(GL_ARRAY_BUFFER, 0)
+        for key in self.vert_data:
+            glDeleteBuffers(1, [self.vert_buffer[key]])
+            glDeleteBuffers(1, [self.norm_buffer[key]])
+
+        self.vert_buffer = {}
+        self.vert_data = {}
+    
+        self.norm_buffer = {}
+        self.norm_data = {}
+
+        self.render_texture_mat = {}
+
+        self.vertex_dim = {}
+        self.n_vertices = {}
+
+    def draw(self):
+        self.draw_init()
+
+        glEnable(GL_MULTISAMPLE)
+
+        glUseProgram(self.program)
+        glUniformMatrix4fv(self.model_mat_unif, 1, GL_FALSE, self.model_view_matrix.transpose())
+        glUniformMatrix4fv(self.persp_mat_unif, 1, GL_FALSE, self.projection_matrix.transpose())
+
+        for mat in self.vert_buffer:
+            # Handle vertex buffer
+            glBindBuffer(GL_ARRAY_BUFFER, self.vert_buffer[mat])
+            glEnableVertexAttribArray(0)
+            glVertexAttribPointer(0, self.vertex_dim[mat], GL_DOUBLE, GL_FALSE, 0, None)
+
+            # Handle normal buffer
+            glBindBuffer(GL_ARRAY_BUFFER, self.norm_buffer[mat])
+            glEnableVertexAttribArray(1)
+            glVertexAttribPointer(1, 3, GL_DOUBLE, GL_FALSE, 0, None)
+
+            glDrawArrays(GL_TRIANGLES, 0, self.n_vertices[mat])
+
+            glDisableVertexAttribArray(1)
+            glDisableVertexAttribArray(0)
+
+        glBindBuffer(GL_ARRAY_BUFFER, 0)
+
+        glUseProgram(0)
+
+        glDisable(GL_MULTISAMPLE)
+
+        self.draw_end()

pifuhd/lib/render/gl/normal_render.py

@@ -0,0 +1,76 @@
+'''
+MIT License
+
+Copyright (c) 2019 Shunsuke Saito, Zeng Huang, and Ryota Natsume
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+'''
+import numpy as np
+
+from .framework import *
+from .cam_render import CamRender
+
+
+class NormalRender(CamRender):
+    def __init__(self, width=1600, height=1200, name='Normal Renderer'):
+        CamRender.__init__(self, width, height, name, program_files=['normal.vs', 'normal.fs'])
+
+        self.norm_buffer = glGenBuffers(1)
+
+        self.norm_data = None
+
+    def set_normal_mesh(self, vertices, faces, norms, face_normals):
+        CamRender.set_mesh(self, vertices, faces)
+
+        self.norm_data = norms[face_normals.reshape([-1])]
+
+        glBindBuffer(GL_ARRAY_BUFFER, self.norm_buffer)
+        glBufferData(GL_ARRAY_BUFFER, self.norm_data, GL_STATIC_DRAW)
+
+        glBindBuffer(GL_ARRAY_BUFFER, 0)
+
+    def draw(self):
+        self.draw_init()
+
+        glUseProgram(self.program)
+        glUniformMatrix4fv(self.model_mat_unif, 1, GL_FALSE, self.model_view_matrix.transpose())
+        glUniformMatrix4fv(self.persp_mat_unif, 1, GL_FALSE, self.projection_matrix.transpose())
+
+        # Handle vertex buffer
+        glBindBuffer(GL_ARRAY_BUFFER, self.vertex_buffer)
+
+        glEnableVertexAttribArray(0)
+        glVertexAttribPointer(0, self.vertex_dim, GL_DOUBLE, GL_FALSE, 0, None)
+
+        # Handle normal buffer
+        glBindBuffer(GL_ARRAY_BUFFER, self.norm_buffer)
+
+        glEnableVertexAttribArray(1)
+        glVertexAttribPointer(1, 3, GL_DOUBLE, GL_FALSE, 0, None)
+
+        glDrawArrays(GL_TRIANGLES, 0, self.n_vertices)
+
+        glDisableVertexAttribArray(1)
+        glDisableVertexAttribArray(0)
+
+        glBindBuffer(GL_ARRAY_BUFFER, 0)
+
+        glUseProgram(0)
+
+        self.draw_end()