Skip to main content

EgoBlur

Overview​

EgoBlur models are open source AI models from Meta to preserve privacy by detecting and blurring faces and license plates from images. We provide two FasterRCNN-based detector models.

For recordings captured by Aria Gen2 devices, we provide both face and license plate anonymization models.

  • Face model is fine-tuned on the EgoBlur Gen1 face model with 250k images, 316k bounding boxes
    • 90k images with 292k bounding boxes from non-egocentric data. 160k images with 24k bounding boxes from Aria Gen2 data.
  • License plate model is trained on 622k images, 676k bounding boxes
    • 584k images with 629k bounding boxes from non-egocentric data. 38k images with 47k bounding boxes from Aria Gen1 data.
  • The EgoBlur face and license plate models are approximately 400 MB each and have ~104 million parameters.
  • The models are based on the Faster RCNN model with a ResNext backbone. The models are trained using Meta’s publicly available Detectron2 and Detectron2go libraries.
  • EgoBlur models are only trained to locate the position of faces and license plates of vehicles within color or grayscale images. The models are not used to track or identify individual faces or license plates.
  • EgoBlur tooling is available to apply the models to PNG, JPEG and MP4 files.

Further reading​

Getting Started​

To demonstrate how EgoBlur models can be applied and to support privacy-preserving research we provide:

  • EgoBlur Demo - Python3 tool for working with PNG, JPEG or MP4 files

    • Go to the EgoBlur Readme for detailed instructions
    • Supported environment:
      • Python version: 3.10-3.12
      • Platform: Fedora 43, Ubuntu 20.04, Ubuntu 22.04, Ubuntu 24.04, MacOS-14, MacOS-15
      • CUDA toolkit (optional): version 12.8. If CUDA is unavailable—either because it’s not supported on the platform (e.g., macOS) or not installed—then inference will run on the CPU.

  • EgoBlur VRS Utilities - C++ tool for working with Aria Gen2 VRS files

    • Coming soon