Add support of face alignment

[jdongca2003]

Face alignment plays an important role in facial analysis (e.g. emotion detection, face recognition). MTCNN has already extracted 5 facial landmark points. We can compare those landmark points with reference ones to construct an affine transform for face alignment. Here I use open source code from https://github.com/ZhaoJ9014/face.evoLVe.PyTorch/blob/master/align/align_trans.py

The changes are minimum:

1. in MTCNN , add extract_aligned_face function
2. add two files (align_trans.py and matlab_cp2tform.py under folder utils
3. update dependency opencv-python in setup.py

With the above changes, I tested this feature. In my experiments, the aligned face size is 112 x 112. The accuracy gain is more than 1% in the internal celebrity dataset.

[AGenchev]

@jdongca2003 I think the face alignment should be integrated into the MTCNN.forward() function and switched on with optional align_face argument or such argument in the MTCNN constructor.
I also don't see the "prewhiten" function used, though I would call it "normalize" or something like it. It might change the model output.

[jdongca2003]

Hi Angel,
Sorry for delay. I agreed with your design (aligned as an argument in forward function). (align_trans.py) is from other repository though I documented it in header. This may not be good. You can add feature request for face align features. Hope that faceNet community developers can pick it up.

Thanks

Jianxiong

[arruw]

I have share simple alignment script here: #34 (comment)
It is not optimal, but it is simple and readable 😄