Under-reconstruction of fast movements

[ChemrovKirill]

Good day and thanks a lot for publishing the code of your great approach!

The results for slow dynamics are awesome even out-of-the-box. However, there is under-reconstruction for fast movements. For example, here is a render of skin gaussians for your dataset with nunchaku. Have you managed to improve the results by tuning hyperparameters (e.g., more joints, less rigid regularization, etc.) or in some other way?

[Guochch]

Thanks for your interest and for sharing your observations! For fast motions like the nunchaku case, you may try the following adjustments in arguments/__init__.py:

• Increase the stage-1 learning rates: position_lr_init_t1 and position_lr_final_t1.

• Increase the number of joint Gaussians with joint_gs_num, which helps stabilize the graph along the nunchaku chain.

• Adjust the ARAP regularization weight (alpha_rigid in LossParamsS1). A higher value makes the model behave more rigidly.

[ChemrovKirill]

Thanks for the answer!
I increased:

• position_lr_init_t1 and position_lr_final_t1 by 2 times
• joint_gs_num by 2 times (to 30_000)
• alpha_rigid by 2 times
  The result became worse) The 50th frame is shown below.
  Perhaps you suggested changes of a different order or something else should be adjusted.

default parameters

adjusted parameters

[Guochch]

Thanks for the feedback!
The nunchuck chain in this sequence is very thin, which makes parameter tuning particularly challenging. This test was done very early on, so the exact implementation and parameters from that time are no longer available.
I’ll help you tune the parameters when I have some time, and get back to you once it’s working better. Thanks for your support! 😊

[ChemrovKirill]

Alright, thanks anyway for the quick responses!