search

SangbumChoi / MobileHumanPose

This repo is official PyTorch implementation of MobileHumanPose: Toward real-time 3D human pose estimation in mobile devices(CVPRW 2021).
MIT License
232 stars 29 forks source link

XS, S, L model settings #37

Open SeheeKim618 opened 1 year ago

SeheeKim618 commented 1 year ago

Hi,

Can you let us know the exact settings for XS, S, L models? I couldn’t reach to the performance of your paper.

Thanks

SangbumChoi commented 1 year ago

@SeheeKim618 https://github.com/SangbumChoi/MobileHumanPose/blob/master/main/config.py You can see this file and as I mentioned in the paper XS has half dimension of small model of voxel parameters

SeheeKim618 commented 1 year ago

image (this is your code)

Thanks for the reply.

By the way, although I set the config.py just like the paper, I was not able to match the performance with that of the paper. (batch size or other settings were the same.) Since there were some unmatched parameters of the config.py with those of the paper, I ran the code after referring to other issues and changing them. Maybe other differences in your code? So I wonder what model the code is built for.

SangbumChoi commented 1 year ago

https://github.com/SangbumChoi/MobileHumanPose/issues/10

As I mentioned in this issue. I used 2 RTX GPUs and It took 3 days. Maybe try different combination of 3d, and 2d dataset. (This is important for performance). However I also remember that my performance had some variation less than 2 MPJPE