Gradient stuck when optimizing smpl parameters

[xyyeah]

I tried to estimate the spml parameters of a real scene based on the colab demo, and the gradient got stuck after 3 to 5 iterations. Camera parameters are estimated based on calibration app (error < 1.5 pixel), and keypoints are estimated based on openpose. From the visualization results, we can see that keypoints2d is correct, any ideas about why the gradient stucks?

s3d 0.340 reg_shapes 0.000
s3d 0.321 reg_shapes 0.001
s3d 0.198 reg_shapes 0.036
s3d 0.159 reg_shapes 0.067
s3d 0.158 reg_shapes 0.067
s3d 0.158 reg_shapes 0.067
s3d 0.158 reg_shapes 0.067
s3d 0.158 reg_shapes 0.067
s3d 0.158 reg_shapes 0.067
s3d 0.158 reg_shapes 0.067
s3d 0.158 reg_shapes 0.067

[chingswy]

Did the results seem to be normal?

[xyyeah]

no, the esitmated smpl is indeed smaller than the person shown in the image. I think the problem may be the result of camera calibration. Here is our result about displaying the cube in the calibration result. It is much worse than yours. Any ideas for improvement? [cid:78e8836c-4cdd-463c-904c-7c3acf14c115]

---

From: Qing Shuai @.> Sent: Sunday, 21 November 2021 9:54 AM To: zju3dv/EasyMocap @.> Cc: xyyeah @.>; Author @.> Subject: Re: [zju3dv/EasyMocap] Gradient stuck when optimizing smpl parameters (Issue #104)

Did the results seem to be normal?

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHubhttps://github.com/zju3dv/EasyMocap/issues/104#issuecomment-974741258, or unsubscribehttps://github.com/notifications/unsubscribe-auth/AVUMZFZWDYAHUMROP6JDERDUNBGMHANCNFSM5ID2VLYA. Triage notifications on the go with GitHub Mobile for iOShttps://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675 or Androidhttps://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub.

[chingswy]

Sorry I can't see your results.

[xyyeah]

sorry, here is the calibration resul

[chingswy]

The calibration result seems to be wrong. You should check the calibration step.

[xyyeah]

Yes, the calibration has been fixed, and the normal smpl model can be obtained at present, but there is still a problem in the repro smpl, the same as #100, what may be the problem?

[chingswy]

I don't know now. If the smpl is right, it's OK.

[xyyeah]

OK, thanks for your quick reply and this exciting project, will close the issue.