Wrong regression result with custom data.

[xiezhy6]

Hi, when I try to use ProHMR to test on my own data, the model fails to regress the correct pose, leading to a terrible fitting result. However, when I test another image with a similar pose, the model works well. (Success case and failure case are shown below.). My question is why the performance of ProHMR varies a lot with similar input images? And is there any solution to fix this issue (e.g., padding/cropping/resizing the input images, modifying the fitting code, etc.)?

Failure case: (from top to down: input image, regression result, fitting result)

Success case:

[xiezhy6]

Hi, I would like to provide some of my test data, in which ProHMR does not work well. Maybe you could test them by yourself and help me analyze the reason for the failure cases. Thanks a lot! :-D

[nkolot]

Is there any specific reason the head of the people is cropped? Is it to preserve privacy? This could create a domain gap between your testing scenario and the images the network was trained on. In such a case you could try to finetuning ProHMR by simultating such types of truncations. Also does OpenPose work well on all those examples or not? I'm curious too see how the OpenPose keypoints look like.