statho / ScoreHMR

ScoreHMR: Score-Guided Diffusion for 3D Human Recovery (CVPR 2024)
MIT License
367 stars 24 forks source link

Smoothing doesn't work #14

Closed tsoydari closed 2 months ago

tsoydari commented 2 months ago

Thanks for your work! Very interesting article. But when testing the provided weights, there is no difference between the results of 4D-Humans and ScoreHMR. The shaking does not go away, especially in the feet.

https://github.com/statho/ScoreHMR/assets/72319918/6285ad29-1537-401e-a15b-35255607a403

On the attached video there are, in order: 4D-Humans, ScoreHMR with default parameters and 1000 iterations of ScoreHMR. Why is there no difference?

statho commented 2 months ago

You can increase smoothing by increasing the W_SMOOTH value here.

In the demo, code we use keypoint detections from ViTPose, which do not detect keypoints on the feet and hands of a person. You could use keypoints from OpenPose if you'd like the model to fit better on the hands and feet.

Sun-Happy-YKX commented 1 month ago

You can increase smoothing by increasing the W_SMOOTH value here.

In the demo, code we use keypoint detections from ViTPose, which do not detect keypoints on the feet and hands of a person. You could use keypoints from OpenPose if you'd like the model to fit better on the hands and feet.

Hello, could you share me the reasonable value range of W_SMOOTH?

statho commented 1 month ago

You can try values between 10 and 70 for W_SMOOTH.