search

Kebii / R2ET

(CVPR 2023) Skinned Motion Retargeting with Residual Perception of Motion Semantics & Geometry
MIT License
174 stars 12 forks source link

The retargeted motion seems to be jittery, is it normal? #9

Open jhonsonlaid opened 1 year ago

jhonsonlaid commented 1 year ago

I followed the README instructions to do inference by running the command: 

python3 inference_bvh.py --config ./config/inference_bvh_cfg.yaml

However, the output I obtained was quite jittery, unlike the smooth motion in the attached demo gif. I tried running the code on both Windows (with Cuda 10.2, Python 3.9, and torch 1.10) and Linux (with Cuda 10.2, Python 3.9, and torch 1.1.0.0).

I am not sure if this jitteriness is due to a limitation in the work or if I made a mistake in my setup. Here is the link to the output generated by the command:

Ortiz_to_Sporty_Granny_Gangnam Style.bvh video

Any suggestions or insights would be greatly appreciated.

Update: When the balancing gate k is set to 0, the jitterness are gone. I think k will introduce jitterness issues.

guolinjie007 commented 1 year ago

+1

miibotree commented 1 year ago

balancing gate k is set to 0, the jitterness are gone. I think k will introduce jitterness issues.

Set k=0 means shape-aware delta_qg is not used. Which means jitter is introduced by shape-aware module. Need to set small k?