About the best model about face-alignment

[jwyang]

@lyttonhao , the model is trained on the trainset of afw, lfpw and helen. To run the model succesffuly, you may need to copy parameter settings from 'models/config_te_best.txt' to config_te.m.

[lyttonhao]

Thanks @jwyang.

[lyttonhao]

Hi, @jwyang, in your implementation, the face detection boxes are obtained using ground truth shapes. I think it's not fair to compare results with the original paper. Or does the original paper also use this strategy as this is consensus in face alignment?

[jwyang]

@lyttonhao ，

In the original paper, I did not find any clues about the face detection strategy. In the time I implement the codes, I was not able to find any provided face rectangles for those dataset. According to the participants of the 300-W chalenge, the initial rectangle is the bounding box of face landmarks.

[jwyang]

@lyttonhao ,

if you get any further information about the strategy for providing initial face rectangles, please let me know. I will appreciate much.

[lyttonhao]

Okay, thanks. @jwyang

[lyttonhao]

@jwyang, I think the training parameters are different from the testing parameters setting in the file 'models/config_te_best.txt'. Right? Can you give the training parameters setting? Thanks.

[jwyang]

@lyttonhao , you can set the training parameters as the testing parameters, which is one of the suggested configuration for this project. Acutally, you can set any parameters as you like, e.g., analogy to the configuration in the original paper.