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

dense pose #8

Closed sankalp2K closed 1 year ago

sankalp2K commented 2 years ago

which dense pose model is used?

koo616 commented 2 years ago

@sankalp2K We use "densepose_rcnn_R_50_FPN_s1x" model in the detectron2 repository :)

sankalp2K commented 2 years ago

Thank You, but how to get segmentation visualization from a dense pose without having the segmentation map blended on top of the original image?

Gzzgz commented 2 years ago

alpha=1

RAJA-PARIKSHAT commented 2 years ago

@Gzzgz what do you mean by alpha=1? I am able to run and get the segmentation map but it is blended with the original image

Arnavgoyal00 commented 2 years ago

How are you able to use detectron to get the densepose image , as in the training dataset?

RAJA-PARIKSHAT commented 2 years ago

@Arnavgoyal00 go to detectron2 repo, clone it and follow the link https://github.com/facebookresearch/detectron2/blob/main/projects/DensePose/doc/GETTING_STARTED.md there will be installation instruction.

Once you have installed you can use https://github.com/facebookresearch/detectron2/blob/main/projects/DensePose/doc/TOOL_APPLY_NET.md#:~:text=python%20apply_net.py%20show%20configs/densepose_rcnn_R_50_FPN_s1x.yaml%20%5C%0Ahttps%3A//dl.fbaipublicfiles.com/densepose/densepose_rcnn_R_50_FPN_s1x/165712039/model_final_162be9.pkl%20%5C%0Aimage.jpg%20bbox%2Cdp_segm%20%2Dv without bbox argument to generate densepose image. But the problem will be it will be blending original image.

For that you will have to change some parameters in visualization. goto Densepose->densepose->vis->densepose_results.py and in class DensePoseResultsFineSegmentationVisualizer set alpha =1

vikashranjan commented 1 year ago

@RAJA-PARIKSHAT I tried setting alpha = 1, but the output is still blended with the original Image.

koo616 commented 1 year ago

Unfortunately, even I have difficulty reproducing densepose images exactly on the dataset. I am currently working on a new dataset using the densepose image obtained by DumpAction in 'apply_net.py'. It seems that there is no significant problem with performance of HR-VITON :)