Face Detection about ai-platform HOT 1 CLOSED

I've cleaned the inference code of one of the SOTA detector implementations (RetinaFace)

https://github.com/hcl14/retinaface-pytorch-inference

It's not the most powerful of the models (MobileNet, the most powerful is ResNet152), however it is possible to port ResNet model to pytorch as well if all the layers can be replicated (see how the weights from mxnet are loaded into pytorch in the original project from which I've forked this: https://github.com/bogireddytejareddy/retinaface-pytorch-inference/blob/master/retinaface.py).

For usage see inference.py.

I removed cython code which made inference faster, but needed manual compiling. I would appreciate if one will pack this into mlflow environment to allow multiple inferences without just relaunching script for every input.

TODO: More powerful ResNet152 model, training code.

THIS is just detector; for face features there seems to be a separate part (this part of retinaface is called insightface? I don't clearly understand), https://github.com/belskikh/insightface-just-works In this repository older MTCNN is used as detector, and insightface as feature extractor. We can later combine this one with our retinaface detector.

SOTA benchmark:

https://paperswithcode.com/task/face-detection

insightface/retinaface (?):

https://paperswithcode.com/paper/190500641#code

from ai-platform.