facebookresearch / sapiens

High-resolution models for human tasks.
https://about.meta.com/realitylabs/codecavatars/sapiens/
Other
4.35k stars 236 forks source link

Huggin Face Depthmap Demo, please consider pulling this until users are able to get consistent depthmaps #64

Closed jones873 closed 1 month ago

jones873 commented 1 month ago

Tried again with the new update, the same problem occurs, and on the bas-relief image nothing was produced. I cannot believe a company as big as Meta would allow something as bad a this to be put out in the public domain....

Untitled Untitled 2

rawalkhirodkar commented 1 month ago

@jones873 This is a failure of the foreground segmentation model - the images you shared are out of distribution of our segmentation labels.

To resolve this, we have added a "no-bg-removal" option to Background Removal Model. This will visualize the depth/normal prediction for the entire image. Note, our models are supervised only on human pixels.

Results on your images below:

image image image image
GloryyrolG commented 2 weeks ago

@jones873 This is a failure of the foreground segmentation model - the images you shared are out of distribution of our segmentation labels.

To resolve this, we have added a "no-bg-removal" option to Background Removal Model. This will visualize the depth/normal prediction for the entire image. Note, our models are supervised only on human pixels.

Results on your images below: image image

image image

hi @rawalkhirodkar then may i ask why not to use a person detector like detectron2/yolox to replace sensitive segmentation? just like what is performed for topdown pose est. thx & best,