search

sail-sg / mvp

NeurIPS-2021: Direct Multi-view Multi-person 3D Human Pose Estimation
Apache License 2.0
325 stars 34 forks source link

Questions about batch size larger than 1 #20

Closed wangjiongw closed 1 year ago

wangjiongw commented 2 years ago

When I run the validation code with given pretrained checkpoint on Panoptic dataset, I found that there is something wrong with batch size larger than 1. For example, When batch size=1, the precision can be reproduced. image

When batch size=2, it seems that the model fails to predict correctly. image

Does anyone get the similar problem? Or can you please give some advice about the reason of this problem? thx!

axhiao commented 2 years ago

Also have same problem.

liqikai9 commented 1 year ago

I got the same problem as well when I reimplement the algorithm. The training speed is really slow and the performance is really bad.

Does anyone have a solution?

wangjiongw commented 1 year ago

In this repo, I didn't find an available solution. But there is an re-implementation of MvP in Mocap (https://github.com/openxrlab/xrmocap), batch size larger than 1 is okay in that repo.

twangnh commented 1 year ago

please refer to https://github.com/openxrlab/xrmocap for improved implementation