tegusi
commented
2 years ago It is mostly because we use different face alignment methods. The face alignment in our framework is not public due to copyright issues. You could preprocess the alignment matrices by OpenCV to see if the results improved.
Wu-ZJ
Thanks for your great work and kind sharing.
I try to reproduce your work on CelebAMask dataset, 24,183 images were selected as the training set and 2834 as the test set. Then I generated label images, edge images and txt files.
After that, I used the code you provided for training. I did not modify the training parameters, but after the training, I found that my training results were nowhere near as good as yours:
Pixel accuracy: 88.680001 Mean accuracy: 62.602909 Mean IU: 52.106180 f1_bg=0.9206227882646921 f1_skin=0.9063355322233573 f1_nose=0.8564767842156585 f1_eye_g=0.7503944737085695 f1_le=0.5846867698606126 f1_re=0.6236390541383516 f1_lb=0.5334376035132539 f1_rb=0.5940682964151555 f1_l_ear=0.5788707588593692 f1_r_ear=0.6270729442117574 f1_imouth=0.6891172758840716 f1_ulip=0.6324672234653816 f1_llip=0.7040783923686705 f1_hair=0.9064766548039499 f1_hat=0.7394414853009562 f1_ear_r=0.23874973724403548 f1_neck_l=0.0 f1_neck=0.812055395815203 f1_cloth=0.7392161008030024 f1_eyes=0.6047181516247372 f1_brows=0.5652373988652979 f1_mouth=0.8338039315469467 f1_overall=0.7834263360796998
The loss in my training have fluctuated between 1.5 and 2.1. And I checked the data set to make sure it had no errors. During the training process, I directly copied your open source code and trained it on train.py. I would like to consult you whether I neglected some key steps? And how can I achieve your results as best as possible?
I would be very grateful if you could reply to me.