Vious / LBAM_Pytorch

Pytorch re-implementation of Paper: Image Inpainting with Learnable Bidirectional Attention Maps (ICCV 2019)

Home Page:http://openaccess.thecvf.com/content_ICCV_2019/papers/Xie_Image_Inpainting_With_Learnable_Bidirectional_Attention_Maps_ICCV_2019_paper.pdf

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Severe artifacts for central hole

C-H-D opened this issue · comments

commented

@Vious Thanks for your excellent work! I'm training your model on Paris StreetView dataset with central square hole, but the results all suffer from severe artifacts like these:
350_100
340_1
350_16
I'm training with batchsize=16, do you have any idea why?

Thank you for your interest in our work, we design the network and loss (gan loss) mainly for irregular holes. For a certain kind of holes like central holes, gan loss plays a very important role. Although we use a parallel structure of the discriminator, I observe that when the holes are very large, even the gan loss converges well, the central area still has many artifacts.
In this central hole inpainting task, I suposse using a local gan which the discriminator is designed for only minizing real/fake distrution of the hole area.

commented

Thanks for your timely reply!😀
哈工大nb!