Strange deformation when aligning SKEL to CMU kicking motion.

[hwy1992129]

Strange deformation when aligning SKEL to CMU kicking motion.

[MarilynKeller]

Hi, glad to hear that you manage to run those two codes. Seems like the SKEL fitting does not converge on every frame, I did not have much time yet finetuning the hyperparamers. You can try the following:

• Run the fit on smaller batches, for example 20 frames (default is 3000 frames in parallel). Each batch pose will get initialized by the last frame of the previous batch result, yielding more stable results:
  python your_poses.npz -F -B 20
  This is slower, so for debugging you can add -D, which will run the fit for only the first batch and save the result. Thus you don't need to wait for the whole sequence to be processed.

• try decreasing the learning rate and double the max-iter and num_step here.

• For better debugging, you can identify the frame that did not converge in aitviewer-skel and then watch the optimization for that specific frame (here frame 215 for example, the batch size miuch be larger than this number).
  python your_poses.npz -F -w 215

Please let me know if you find a more stable combination, I'll update the code in consequence.