- Just an integration of face recognition pipeline to pure pytorch version
- References: to be added
- TensorRT inference: to be added
- Many necessary information is missing now (e.g. environment configuration), which will be supplemented later
- input: a 3-channel image
- Face Detection: use MTCNN to detect all possible bounding boxes of 'faces', each with 5 keypoints
- Image Crop: crop all the faces according to bounding boxes
- Keypoint Alignment: use skimage to project detected keypoints to standard location
- [Function] Face Anti-spoofing: a pytorch model to detect whether the cropped face represents real faces or not
- Feature Encoding: in local image dataset, execute feature encoding to transfer image into 128 features, all stored in json format
- [Function] Face Recognition: a pytorch model to recognize the corresponding people's name from given face image
-
Newly created repositoty with newly updated packages, easy to configure the environment
-
Friendly to all users, giving explicit coding style and coding process, which is not only serving for ourselves and just storing it on GitHub
-
Integrating very small but accurate models, all based on GPU so as to make a slightest mobile face recognition pipeline, or even CPU with still fast inference, with which to construct a face-recognition platform in your own PC at any time and any place
-
[to be added] create GUI to help register new people's face, making dataset maintenance easier
-
[to be added] apply TensorRT to accelerate inference in some mobile platforms (e.g. NVIDIA Jetson TX2)
-
[to be added] more complete pipeline, including sex detection, age detection, etc. All successful appliance will just depend on small model(s)
-
[Tested] Now all the models only cover 553MB of space in your GPU.