Dataset Card for Face-Obfuscated-Imagenet

---

ImageNet Usage Agreement

To use the face-obfuscated-imagenet dataset—which includes face-blurred versions of ILSVRC2012 and its subsets you must agree to the official ImageNet Terms of Access.

These terms include (but are not limited to):

• You must be an approved academic or research user.
• Use of the dataset is restricted to non-commercial research and educational purposes.
• Redistribution or external hosting of the underlying image data is strictly prohibited.

How to Gain Access

1. Visit the official ImageNet download page: https://image-net.org/download-images
2. Sign up using your institutional credentials.
3. Read and accept the ImageNet Terms of Access.
4. Submit a request for access to the ILSVRC2012 dataset.
5. Use face-obfuscated-imagenet only after your access is approved.

By downloading or using this dataset, you confirm that you have read, understood, and agreed to the ImageNet terms of use.

---

Dataset Details

Dataset Description

Face-Obfuscated-Imagenet is a privacy-enhanced variant of the ILSVRC2012 ImageNet dataset, featuring face-obfuscated (blurred) images across multiple standard subsets, as well as the original subsets. It supports large-scale image classification research while preserving privacy. This dataset includes the following config options:

• 1k
  Standard ILSVRC2012 with ~1.2 million images across 1000 classes for classification and recognition tasks.

• 100
  Curated subset of 100 ImageNet classes for faster experimentation with diverse categories.

• 10 (Imagenette)
  Lightweight subset of 10 easy-to-classify classes for quick prototyping and educational use. Utilizes validation images in place of test images. Since this uses validation images in place of test images, it is not identical to the original Imagenette dataset. However it contains the same classes and can be used for straightforward testing. Reported results should instead use 1k or 100 (or their blurred counterparts).

• noface-1k
  Face-obfuscated version of the 1k dataset with human faces blurred for privacy-preserving research.

• noface-100
  Face-obfuscated subset matching the 100 classes of ImageNet-100 with blurred faces.

• noface-10
  Face-obfuscated Imagenette subset with 10 classes, protecting identities via face blurring. Utilizes validation images in place of test images. This is the blurred equialent of the 10 mentioned above.

All versions share a cache. When you use one of the versions, the downloads and initial processing for all versions will be completed. In future runs, this shared cache will avoid repetitive downloads.

Dataset Sources

Original ImageNet (ILSVRC2012)

• Homepage: ImageNet Challenge 2012
• Paper: Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., ... & Fei-Fei, L. (2015).
  ImageNet Large Scale Visual Recognition Challenge.
  International Journal of Computer Vision, 115, 211–252.

Face-Obfuscated ImageNet

• Homepage: Face Obfuscation Project
• Paper: Yang, K., Yau, J., Fei-Fei, L., Deng, J., & Russakovsky, O. (2022).
  A Study of Face Obfuscation in ImageNet.
  International Conference on Machine Learning (ICML).

Dataset Structure

This dataset includes the following subsets with face-obfuscated versions available. The numbers show total images with the face-obfuscated images present in (parentheses).

Version	Training Images	Validation Images
1k (noface_1k)	1,281,167 (1,281,066)	50,000 (49,997)
100 (noface_100)	126,689 (126,683)	5,000 (4,999)
10 (noface_10)	12,894 (12,893)	500 (500)

Additional Features

Each example in the dataset includes two boolean features:

• is_one: Boolean flag used in noface training data for subset identification. Always False for validation splits and non-face-blurred datasets. Corresponds to a 1% subset used in semi-supervised learning.
• is_ten: Boolean flag used in noface training data for subset identification. Always False for validation splits and non-face-blurred datasets. Corresponds to a 10% subset used in semi-supervised learning.

These flags are extracted from the metadata format in face-obfuscated training data and can be used for specialized training procedures or subset analysis.

Example Usage

To use a specific version, you need to specify the correct config. The following code shows how all versions could be used.

import datasets
from aiohttp import ClientTimeout

dataset_name = "randall-lab/face-obfuscated-imagenet"
timeout_period = 500000  # intentioanlly long enough so timeout will not ever occur
storage_options = {"client_kwargs": {"timeout": ClientTimeout(total=timeout_period)}}

# List of dataset subsets to load
dataset_names = ["1k", "100", "10", "noface-1k", "noface-100", "noface-10"] # these are the config options

# this will test dataset loading of train and validation splits for each version
for i, name in enumerate(dataset_names, start=1):
    for split in ["train", "validation"]:
        dataset = datasets.load_dataset(
            dataset_name,
            name=name, # this needs to be set as your desired config
            split=split,
            trust_remote_code=True,
            storage_options=storage_options,
        )

        example = dataset[0]  # type: ignore
        image = example["image"]
        label = example["label"]
        is_one = example["is_one"]  # Boolean indicating subset membership (True only for noface training data)
        is_ten = example["is_ten"]  # Boolean indicating subset membership (True only for noface training data)
        print(f"{split.title()} label:", dataset.features["label"].names[label])  # type: ignore
        print(f"is_one: {is_one}, is_ten: {is_ten}")
        image.save(f"output_image_{i}_{split}.jpg", format="JPEG")  # type: ignore

Citation

BibTeX:

@article{russakovsky2015imagenet,
  title={Imagenet large scale visual recognition challenge},
  author={Russakovsky, Olga and Deng, Jia and Su, Hao and Krause, Jonathan and Satheesh, Sanjeev and Ma, Sean and Huang, Zhiheng and Karpathy, Andrej and Khosla, Aditya and Bernstein, Michael and others},
  journal={International journal of computer vision},
  volume={115},
  pages={211--252},
  year={2015},
  publisher={Springer}
}
@inproceedings{yang2022study,
  title={A Study of Face Obfuscation in ImageNet},
  author={Yang, Kaiyu and Yau, Jacqueline and Fei-Fei, Li and Deng, Jia and Russakovsky, Olga},
  booktitle={International Conference on Machine Learning (ICML)},
  year={2022}
}