search

eth-siplab / AvatarPoser

Official Code for ECCV 2022 paper "AvatarPoser: Articulated Full-Body Pose Tracking from Sparse Motion Sensing"
https://siplab.org/projects/AvatarPoser
MIT License
288 stars 49 forks source link

why the prediction is far more unstable than the gt pose? #14

Closed yukichou closed 1 year ago

yukichou commented 1 year ago

I used the pretrained model to predict the test set of AMASS dataset and visualized the prediction, but the output is severe jittering comparing to the GT pose. ( I visualized the prediction and the GT in unity, the GT pose is smooth and accurate) Regarding to the root orientation which comes from the stabilizer, is badly shaking too. BTW, the test score is 1.98/2.51/24.11. Does anyone know this problem?

yukichou commented 1 year ago

Also, I found out the ik_module is not being used in the code, while the framework in the paper is using it to optimize the arm's joints, why is that?

jiaxi-jiang commented 1 year ago

Hi, it seems that I uploaded a wrong pretrained model which was trained on another data split, so please just retrain the model from scratch. Sorry I haven't time to clean the code of IK yet, but it is easy to implement.