Anonymizer

An anonymizer for pravicy protection of human faces and vehicle plates. This is a customized fork originally from archived project of understand-ai. The original code is developed with tensorflow-1.11 which is a fairly old version with poor GPU support in recent GPUs. With some tricks, the ONNX-RUNTIME provides a work around to running on fairly new GPUs (i.e. CUDA 11.*). Guidelines for ONNX-RUNTIME version is here.

You can use this anonymizer to mask your datasets which is selected to publish.

Installation

• Clone the Repo

  git clone https://github.com/fusionportable/Anonymizer.git

• Download the pre-trained weights

  # two .onnx files
  face.onnx
  plate.onnx

• Build Docker Container

  # without ros
  docker build . -t anonymizer:ort-1.14-cuda-11.6

• Install Dependencies

  # tested with CUDA11.8 with onnxruntime=1.14
  cd anonymizer
  pip install --upgrade pip
  pip install -r requirements.txt

Usage

• Anonymize images

python3 anonymizer/bin/anonymize_onnx.py \
--input /absolute/path/to/input/images/folder \
--output /absolute/path/to/output/images/folder \ 
--weights /absolute/path/to/weights/folder \ --image-extensions jpg \ 
---face-threshold 0.3 --plate-threshold 0.3 \
--write-detections (default false)

• Anonymize rosbags

  TODO

TODO:

• Add docker image with CUDA11.6 compatibility.
• Add docker image with ROS1 compatibility.
• Add instructions on easy run.
• Add figures for illustration.
• Add scripts directly process the rosbags.
• Refactor the anonymizer in the same branch that is compatiable to plug in detector and obfuscators.