Introducing EgoBlur, a state-of-the-art obfuscation system developed by Meta as part of their commitment to responsible innovation. EgoBlur is designed to protect privacy while accelerating AI and ML research, and is now available under an open-source license (Apache 2.0) for both commercial and non-commercial use. EgoBlur was first developed for Meta's Project Aria program, and has been used internally since 2020. The system uses advanced face and license plate obfuscation techniques to help maintain privacy while still allowing researchers to access valuable data.

With EgoBlur, researchers can now work on AI and ML projects without compromising the privacy of those around them. This is especially important in applications such as computer vision, where large amounts of data are often required to train models effectively. We believe that EgoBlur will be a valuable tool for the external research community, and we are excited to make it available under an open-source license. By using EgoBlur, researchers can focus on advancing AI and ML technology while also maintaining the trust of the public.

This repository contains demo of EgoBlur models with visualizations.

This code requires conda>=23.1.0 to install dependencies and create a virtual environment to execute the code in. Please follow the instructions here to install Anaconda for your machine.
We list our dependencies in environment.yaml file. To install the dependencies and create the env run:

conda env create --file=environment.yaml

# After installation, check pytorch.
conda activate ego_blur
python
>>> import torch
>>> torch.__version__
'1.12.1'
>>> torch.cuda.is_available()
True

Please note that this code can run on both CPU and GPU but installing both PyTorch and TorchVision with CUDA support is strongly recommended.

First download the zipped models from given links. Then the models can be used as input/s to CLI.

A brief description of CLI args:
--face_model_path use this argument to provide absolute EgoBlur face model file path. You MUST provide either --face_model_path or --lp_model_path or both. If none is provided code will throw a ValueError.
--face_model_score_threshold use this argument to provide face model score threshold to filter out low confidence face detections. The values must be between 0.0 and 1.0, if not provided this defaults to 0.1.
--lp_model_path use this argument to provide absolute EgoBlur license plate file path. You MUST provide either --face_model_path or --lp_model_path or both. If none is provided code will throw a ValueError.
--lp_model_score_threshold use this argument to provide license plate model score threshold to filter out low confidence license plate detections. The values must be between 0.0 and 1.0, if not provided this defaults to 0.1.
--nms_iou_threshold use this argument to provide NMS iou threshold to filter out low confidence overlapping boxes. The values must be between 0.0 and 1.0, if not provided this defaults to 0.3.
--scale_factor_detections use this argument to provide scale detections by the given factor to allow blurring more area. The values can only be positive real numbers eg: 0.9(values < 1) would mean scaling DOWN the predicted blurred region by 10%, whereas as 1.1(values > 1) would mean scaling UP the predicted blurred region by 10%.
--input_image_path use this argument to provide absolute path for the given image on which we want to make detections and perform blurring. You MUST provide either --input_image_path or --input_video_path or both. If none is provided code will throw a ValueError.
--output_image_path use this argument to provide absolute path where we want to store the blurred image. You MUST provide --output_image_path with --input_image_path otherwise code will throw ValueError.
--input_video_path use this argument to provide absolute path for the given video on which we want to make detections and perform blurring. You MUST provide either --input_image_path or --input_video_path or both. If none is provided code will throw a ValueError.
--output_video_path use this argument to provide absolute path where we want to store the blurred video. You MUST provide --output_video_path with --input_video_path otherwise code will throw ValueError.
--output_video_fps use this argument to provide FPS for the output video. The values must be positive integers, if not provided this defaults to 30.

Download the git repo locally and run following commands. Please note that these commands assumes that you have a created a folder /home/${USER}/ego_blur_assets/ where you have extracted the zipped models and have test image in the form of test_image.jpg and a test video in the form of test_video.mp4.

The model is licensed under the Apache 2.0 license.

See contributing and the code of conduct.

If you use EgoBlur in your research, please use the following BibTeX entry.

@misc{raina2023egoblur,
      title={EgoBlur: Responsible Innovation in Aria},
      author={Nikhil Raina and Guruprasad Somasundaram and Kang Zheng and Sagar Miglani and Steve Saarinen and Jeff Meissner and Mark Schwesinger and Luis Pesqueira and Ishita Prasad and Edward Miller and Prince Gupta and Mingfei Yan and Richard Newcombe and Carl Ren and Omkar M Parkhi},
      year={2023},
      eprint={2308.13093},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}