The problem of the edge

[yuzhesu]

Is it because the external cropping loss of the cropping frame is ignored, so the outermost part of the generated picture is not very effective?How to solve it?

[SeitaroShinagawa]

Yes, my current implementation ignored the outside of cropping area due to the pre-trained VGG setting which requires input size of 224.

If you want to apply loss function to all area strictly, I think downsampling or chainer.functions.resize_images may help you.
https://docs.chainer.org/en/stable/reference/generated/chainer.functions.resize_images.html#chainer.functions.resize_images Note that I don't guarantee it works well because I have never used this function.

[yuzhesu]

Are you referring to downsampling or chainer.functions.resize_images at updater.py def vgg_extract?

[SeitaroShinagawa]

Actually not vgg_extract itself, but the input of vgg_extract. It requires 224 as input, so you can resize default input size 256 to 224 using the above ways.

[quelladai]

Do you mean that resize the size of the input images and masks to 224*224?

[SeitaroShinagawa]

Yes, it depends on VGG pre-trained model.