search

sangyun884 / HR-VITON

Official PyTorch implementation for the paper High-Resolution Virtual Try-On with Misalignment and Occlusion-Handled Conditions (ECCV 2022).
831 stars 171 forks source link

Offset Generated During Training #34

Closed Atotem closed 1 year ago

Atotem commented 1 year ago

Hi, I'm training with my own dataset and had some initial results with 1000 epochs. However, I noticed that the targeted garment moves from its position during training. Do you know what this can be happening? I followed the same preprocessing process but can't figure out at this point what can be causing it. Greetings!

00364_00018

Sincerely, @Atotem

koo616 commented 1 year ago

Hello @Atotem. I think you have a problem with your human parse map. Now you can check the detailed preprocess in 'Preprocess.md'. Please check it :)

Atotem commented 1 year ago

Hi @koo616 , Are your sure you used torchgeometry.image.GaussianBlur()? and not torchvision.transforms.GaussianBlur()? Because for torchgeometry I have to transform the image to tensor. Also, I get images but with a lot of unwanted data in the edges. I tried ImageFilter.ModeFilter(size=3) from Pillow but I'm still trying to figure out the kernel size. Any ideas?

khurramkhalil commented 1 year ago

Hi @koo616, I am facing similar issues. I have usedImageFilter.ModeFilter(size=10). Can you share your image agnostic and pose generation code so that we both can reconcile? Thanks