search

JiangWenPL / multiperson

Code repository for the paper: "Coherent Reconstruction of Multiple Humans from a Single Image" in CVPR'20
https://jiangwenpl.github.io/multiperson/
376 stars 50 forks source link

Speed for fitting SMPL-X to 3D key points #8

Closed ZhengyiLuo closed 3 years ago

ZhengyiLuo commented 4 years ago

Hi,

Thanks a ton for the great work!

I am trying out the script to fit SMPL-X to 3D keypoints and are noticing a speed of about 40 secs per frame; This is rather slow even with parallelzation. Any suggestions?

Sicanno commented 3 years ago

Hi, I get the same problem. By the way, how do we get the 'data/h36m/extras/h36m_single_train_openpose.npz', just translate this .pkl file into the .npz? Because it seems that only the pose and shape parameters of 3.6m has been used.

Sicanno commented 3 years ago

Also, the loss of idx 15635 is nan. Maybe it will cause some error.

tiangexiang commented 3 years ago

Hi, do you guys solve the speed problem?

ZhengyiLuo commented 3 years ago

@tiangexiang This repo has the fitted H36M from this method available for download: https://github.com/mks0601/I2L-MeshNet_RELEASE

tiangexiang commented 3 years ago

@ZhengyiLuo Hi, thanks for your response! It seems the SMPL params from https://github.com/mks0601/I2L-MeshNet_RELEASE cannot fit directly to the dataloader of this repo, how did you convert the formats? Or did you use a completely new dataloader for h36m? Thanks again!

GloryyrolG commented 1 year ago

Hi,

Thanks a ton for the great work!

I am trying out the script to fit SMPL-X to 3D keypoints and are noticing a speed of about 40 secs per frame; This is rather slow even with parallelzation. Any suggestions?

Using 8 processes, it takes 400k * 40 / 8 / 3600 / 24 = 23 days...