facebookresearch / detectron2

Detectron2 is a platform for object detection, segmentation and other visual recognition tasks.
https://detectron2.readthedocs.io/en/latest/
Apache License 2.0
30.56k stars 7.49k forks source link

a problem about perdicted uv coordinates #522

Closed uyzhang closed 1 year ago

uyzhang commented 4 years ago

when I reappeared the paper's (https://arxiv.org/pdf/1905.08776.pdf) result, the problem about DensePose appeared. There are the results when I used DensePose to infer. I used the model downloaded from the Model zoo, url(https://dl.fbaipublicfiles.com/densepose/densepose_rcnn_R_101_FPN_s1x/143908726/model_final_ad63b5.pkl)

the images: hd_00_00_10750 hd_00_00_10765

the v coordinate results show: 10750 0001 10765 0001

the textures that get using densepose results and images: ts_2 ts_2(1)

So, why can get contrary v coordinate of chest part in the same view? And, the textures of chest get from uv coordinate and image with the two actions are contrary

uyzhang commented 4 years ago

I just used the old model(https://dl.fbaipublicfiles.com/densepose/DensePose_ResNet101_FPN_s1x-e2e.pkl) in the Model zoo, the problem did not happen again. So the best model(https://dl.fbaipublicfiles.com/densepose/densepose_rcnn_R_101_FPN_s1x/143908726/model_final_ad63b5.pkl) has some problems?

vkhalidov commented 4 years ago

Dear @xduzy16, the problem with flipped body coordinates has recently been fixed in 5ac96b2, the model zoo will be updated soon.

uyzhang commented 4 years ago

thanks!

vkhalidov commented 4 years ago

the bug has been fixed in 5ac96b2 the model zoo has been updated on Feb 26 in 7c25819 body coordinates should not be flipped any more