Training-set alpha still has big holes #283
Description
I visualized some validation samples: from left to right we have the source image, predicted alpha (pred_pha) and ground-truth alpha (true_pha).
All samples were drawn from the training split, yet the predictions still show obvious missing regions.
Is this normal? What are the common remedies?
