Dataset Card for ImageNet-Sketch

Dataset Summary

ImageNet-Sketch data set consists of 50000 images, 50 images for each of the 1000 ImageNet classes. We construct the data set with Google Image queries "sketch of __", where __ is the standard class name. We only search within the "black and white" color scheme. We initially query 100 images for every class, and then manually clean the pulled images by deleting the irrelevant images and images that are for similar but different classes. For some classes, there are less than 50 images after manually cleaning, and then we augment the data set by flipping and rotating the images.

The scripts used to conduct queries and clean images can be found in the GitHub repository.

Supported Tasks and Leaderboards

• image_classification: The goal of this task is to classify a given image into one of 1000 ImageNet classes. The leaderboard is available here.

The goal of the leaderboard is to evaluate the out-of-domain classification performance of vision models trained on ImageNet. The evaluation metrics used in the leaderboard are top-1 accuracy and top-5 accuracy.

Languages

The class labels in the dataset are in English.

Dataset Structure

Data Instances

A sample from the training set is provided below:

{
    'image': <PIL.JpegImagePlugin.JpegImageFile image mode=RGB size=400x530 at 0x7FB2EF5D4A90>,
    'label': 320
}

Data Fields

The data instances have the following fields:

• image: A PIL.Image.Image object containing the image. Note that when accessing the image column: dataset[0]["image"] the image file is automatically decoded. Decoding of a large number of image files might take a significant amount of time. Thus it is important to first query the sample index before the "image" column, i.e. dataset[0]["image"] should always be preferred over dataset["image"][0].
• label: an int classification label.

The labels are indexed based on a sorted list of synset ids such as n07565083 which we automatically map to original class names. The original dataset is divided into folders based on these synset ids. To get a mapping from original synset names, use the file LOC_synset_mapping.txt available on Kaggle challenge page. You can also use dataset_instance.features["label"].int2str function to get the class for a particular label index.

Data Splits

Dataset Creation

Curation Rationale

From the paper:

Inspired by the Sketch data of (Li et al., 2017a) with seven classes, and several other Sketch datasets, such as the Sketchy dataset (Sangkloy et al., 2016) with 125 classes and the Quick Draw! dataset (QuickDraw, 2018) with 345 classes, and motivated by absence of a large-scale sketch dataset fitting the shape and size of popular image classification benchmarks, we construct the ImageNet-Sketch data set for evaluating the out-of-domain classification performance of vision models trained on ImageNet.

Source Data

Initial Data Collection and Normalization

The initial data collection and normalization is inherited from ImageNet. More information on it can be found here.

Additional preprocessing from the paper:

We construct the data set with Google Image queries “sketch of ”, where is the standard class name. We only search within the “black and white” color scheme. We initially query 100 images for every class, and then manually clean the pulled images by deleting the irrelevant images and images that are for similar but different classes. For some classes, there are less than 50 images after manually cleaning, and then we augment the data set by flipping and rotating the images.

Who are the source language producers?

The source language is inherited from ImageNet. More information on the source language produces can be found here.

Annotations

Annotation process

The annotations are inherited from ImageNet. More information about the process can be found here.

Who are the annotators?

The same as in ImageNet.

Personal and Sensitive Information

[More Information Needed]

Considerations for Using the Data

Social Impact of Dataset

[More Information Needed]

Discussion of Biases

The biases are inherited from ImageNet. More information about the process can be found here.

Other Known Limitations

1. Since most of the images were collected from internet, keep in mind that some images in ImageNet-Sketch might be subject to copyrights.

Additional Information

Dataset Curators

Authors of Learning Robust Global Representations by Penalizing Local Predictive Power:

• Haohan Wang
• Songwei Ge
• Eric P. Xing
• Zachary C. Lipton

The dataset was curated using the scripts found in the GitHub repository.

Licensing Information

[More Information Needed]

Citation Information

@inproceedings{wang2019learning,
        title={Learning Robust Global Representations by Penalizing Local Predictive Power},
        author={Wang, Haohan and Ge, Songwei and Lipton, Zachary and Xing, Eric P},
        booktitle={Advances in Neural Information Processing Systems},
        pages={10506--10518},
        year={2019}
}

Contributions

Thanks to @nateraw for adding this dataset.