DL3DV/DL3DV-Benchmark

DL3DV Benchmark Download Instructions

This repo contains all the benchmark data, including a README, License, colmaps/images (compatible to nerfstudio and 3D gaussian splatting), scene labels and the performances of methods reported in the paper (ZipNeRF, 3D GS, MipNeRF-360, nerfacto, Instant-NGP).

Download

As the whole benchmark dataset is very big (~2.1T), we provide two ways to download: full benchmark dataset download or use a script to download a subset for memory sensitive cases.

Full benchmark dataset download

If you have enough space (more than 2.1T), download the full benchmark is simple:

# Make sure you have git-lfs installed
# (https://git-lfs.github.com/)
git lfs install

git clone https://huggingface.co/datasets/DL3DV/DL3DV-10K-Benchmark

Script download

Sometimes you may just need to flexibly download a subset the benchmark, e.g. just download several scenes, or just need images with 960P resolution (images_4 level used in the paper). To provide this flexibiliy, we provide a download.py script for use. Use this link to download.

This download script provies several different options to use:

• Download the full dataset (which is equivalent to git clone method). In total 2.1T.
• Download the full dataset with only 960P images. In total 100~150G.
• Download with specific scene name (hash name)

Environment Setup

The download script relies on huggingface hub, tqdm, and pandas. You can download by the following command in your python environment. The download script was

pip install huggingface_hub tqdm pandas

After downloading huggingface_hub, remember to login first to get ready for download.

# in terminal, use the following command and your huggingface token to login
huggingface-cli login

Download the full benchmark

To download the full dataset, use this command:

# Note, it is suggested to use --clean_cache flag as it saves space by cleaning the cache folder created by huggingface hub API. 
python download.py  --subset full --clean_cache

Download the full benchmark with 960P resolution (same with the paper)

Not all the methods can handle multi-resolution. Some methods have assumptions on the input resolution. So the paper uses 960P.

# Note, it is suggested to use --clean_cache flag as it saves space by cleaning the cache folder created by huggingface hub API. 
python download.py  --subset full --only_level4 --clean_cache

Download with specific scene name (hash name)

There is a benchmark preview page in https://github.com/DL3DV-10K/Dataset. If you just need a specific hash (e.g. 0853979305f7ecb80bd8fc2c8df916410d471ef04ed5f1a64e9651baa41d7695), use the following command:

# Note, it is suggested to use --clean_cache flag as it saves space by cleaning the cache folder created by huggingface hub API. 
# e.g. a scene with hash 0853979305f7ecb80bd8fc2c8df916410d471ef04ed5f1a64e9651baa41d7695
python download.py  --subset hash --hash 0853979305f7ecb80bd8fc2c8df916410d471ef04ed5f1a64e9651baa41d7695 --only_level4