When dct_bool[img_idx] == 0, the pixel values in img_train_SFM[img_idx,0,:,:] will be all zeros, is this training data correct?

[hellozhaozheng]

but when dct_bool[img_idx] == 0, the pixel values in img_train_SFM[img_idx,0,:,:] will be all zeros, is this training data correct?

Originally posted by @hellozhaozheng in https://github.com/majedelhelou/SFM/issues/3#issuecomment-682281828

[majedelhelou]

In general, we noticed empirically that applying SFM on 50% of the training data, while the rest is augmented by only noise maps, provides better generalization strength to unseen noise levels. The networks can learn the noise distribution disentangled from the signal. However, if you are only interested in performance on training noise levels, you can apply SFM on 50% of the training data, and keep the rest as unmasked​ (original image) data. We observed that this can improve our results (better than in our paper) on the training noise levels [10,50], sometimes significantly.