caizhongang / SMPLer-X

Official Code for "SMPLer-X: Scaling Up Expressive Human Pose and Shape Estimation"
https://caizhongang.github.io/projects/SMPLer-X/
Other
1.01k stars 73 forks source link

Problem regarding estimation accuracy in Foot Region #45

Closed Marlod390 closed 7 months ago

Marlod390 commented 10 months ago

Hi, I ran the project on other videos and it worked generally well. But I observed some differences between the pose estimation output and the actual human pose, especially in the foot area. The attached image shows a clear divergence, where the estimated pose points are not aligned with the physical position of the foot:

000001 000200

I tried models s to h and the same problem occurred in all models. Can you give your opinion on this matter? Is this a known issue or could you recommend any method to improve the performance?

Thanks in advance.

Wei-Chen-hub commented 9 months ago

In my opinion, it looks like a projection error (i.e. scale), the knee & leg also seems to be smaller that the real size.

Please upload the original video so that i can check it out from my side.

Marlod390 commented 9 months ago

Thanks for your reply. Here is the original video: https://github.com/caizhongang/SMPLer-X/assets/67206431/96a1471e-b17d-401c-bf91-3817cfba144c

For some reason this link does not work on Firefox, please use chrome to download.

Wei-Chen-hub commented 7 months ago

Apologize for the late response, I have done a throughly check, and found a bug in 3DPW dataset loader. There was a mismatch of bbox size (1 vs. 1.25) during image preprocessing, therefore 2d keypoints were smaller than actual. The affect is that model would yield smaller person on 3DPW-like data.

3219

We have fixed this issue and re-done the training. With updated model, the problem is solved.

smplerx_issue45_fix

Please try the updated model, hope this reply would help.

Wei-Chen-hub commented 7 months ago

I will close this issue for now, if you encounter any other situations, feel free to raise a new issue.