Hello @shaoanlu, first of all thank you for a great job porting the models to keras.
I'm unsure of what exactly is this step doing. I think that corresponds to a center crop of a tensor with shape (b,128,128,3) resulting in a tensor with shape (b,112,112,3), but the ir50 preprocessing layer is resizing the tensor to 112 x 112 beforehand.
Could this be a mistake?
Edit: I can confirm this shouldn't be in the preprocessing step, at least in combination with alignment.
Evaluation on LFW using face-alignment with MTCNN landmarks increased from 99.62% to 99.82% after removing that step.
Hello @shaoanlu, first of all thank you for a great job porting the models to keras.
I'm unsure of what exactly is this step doing. I think that corresponds to a center crop of a tensor with shape (b,128,128,3) resulting in a tensor with shape (b,112,112,3), but the ir50 preprocessing layer is resizing the tensor to 112 x 112 beforehand.
Could this be a mistake?
Edit: I can confirm this shouldn't be in the preprocessing step, at least in combination with alignment.
Evaluation on LFW using face-alignment with MTCNN landmarks increased from 99.62% to 99.82% after removing that step.