About my own mask

[JinMengKang]

Hello, dear author. I want to test with my own mask, during the training phase, I set: mask_type='random' fix_mask=false But the result is very poor. Can you help me find out why? This is the poor result,The white place is my own mask.

[Zhaoyi-Yan]

Hmm, I think several problems may exist:

1. Your dataset is too small. You can use 'RandomResizeCrop', 'flipping', etc as your data augmentation methods.
2. You should adjust https://github.com/Zhaoyi-Yan/Shift-Net/blob/master/train.lua#L102 local res to larger value, so that you can get smaller continuous mask during training.(Make the area of mask almost fit that of your real mask).
3. https://github.com/Zhaoyi-Yan/Shift-Net/blob/master/train.lua#L57 threshold is, in fact, just to define the masked region in feature space, you may just resize the whole mask using nearest downsample.
4. You can add an easy extension on my work, just to more adaptively combine F_{shift} with F_{enc} and F_{dec}. My work is just an opening point to combine exemplar-based inpainting with CNN inpainting, we just stack these features directly and let CNN select useful information. You can think up an idea to combine these features more adaptively.