Bad results for large masks

[prerna-grg]

I tried testing your code on masks of large sizes and the output is quite bad. Any idea for some workaround that can be done for this?

[shepnerd]

1. Using a network with more capacity and training it with a large size mask and more time.
2. If you wanna recover less semantically sensitive regions (e.g., grassland, sky, and rocks), you can consider applying contextual attention module or non-local module to the bottleneck of the generator. They work well for even large missing regions.
3. If you are using masks generated by stokes, sparse convolution (or partial convolution) and gated convolution will be helpful. Meanwhile, you can consider removing adversarial training and replace l1 loss with perceptual loss.

Prerna Garg notifications@github.com 于2019年4月1日周一 下午3:55写道：

I tried testing your code on masks of large sizes and the output is quite bad. Any idea for some workaround that can be done for this?

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/shepnerd/inpainting_gmcnn/issues/5, or mute the thread https://github.com/notifications/unsubscribe-auth/AE1PCAoP_xcUNHx7zzTXdovgjCx3PcAtks5vcbuIgaJpZM4cVE-c .