search

cuiaiyu / dressing-in-order

(ICCV'21) Official code of "Dressing in Order: Recurrent Person Image Generation for Pose Transfer, Virtual Try-on and Outfit Editing" by Aiyu Cui, Daniel McKee and Svetlana Lazebnik
https://cuiaiyu.github.io/dressing-in-order
Other
507 stars 126 forks source link

Testing with custom images of garment only #50

Closed dain5832 closed 1 year ago

dain5832 commented 2 years ago

First of all, thanks for the amazing project!

My teammates and I are using your work in part of our course project, and we are using the FashionIQ dataset, which is mostly consisted of garment-only images, for inference on your pretrained model Can you give us a more detailed explanation of how to preprocess the dataset and run the inference in the case of garment-only images?

We found you mentioned that P_{g_k} is going to be an empty heatmap when the image is garment only, but I'm not really sure how can we implement it. And it seems like the parser does not work well on garment-only images, so if there are any other methods we can try to parse the garment from the background.

Also, we are struggling with the face overfitting issue and would like to ask if using garment-only images will help resolve the issue since the pose is not transferred.

snaiws commented 1 year ago

face overfitting issue : #32

garment-only image issue : I'm wondering as well

snaiws commented 1 year ago

hey i solved garment-only image issue

treat garment as human model but set all keypoints as -1 and instead of human parsing, use image segmentation. there would be a garment area(lip label) and background(0) here's example jacket3

dain5832 commented 1 year ago

Thanks!! I'll check that out

netsalesbg-ltd commented 1 year ago

@snaiws Do you have to retrain the whole network again for garment only? Does retraining fix face overfitting?