Giters

yz-cnsdqz / PSI-release

official implementation of CVPR'20 oral paper: Generating 3D People in Scenes without People.: https://ps.is.tuebingen.mpg.de/publications/smpl-x-conditional-vae-prox-scene-constraints

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

3D human pose estimation

sulei1998 opened this issue · comments

commented

Hello! Nice work! But i have a question, when i read your paper, i notice that you said your work can improve 3D pose estimation from monocular images. So i want to know how to combine your work with the prox, is there any codes? I have tested the [PROX] (https://github.com/MohameHassan/prox) before and i find that prox fit a image cost lots of time. So I'm curious if your method can increase the running speed?

Thanks a lot!

commented

Many thanks for interests. Answers are below:

Hello! Nice work! But i have a question, when i read your paper, i notice that you said your work can improve 3D pose estimation from monocular images. So i want to know how to combine your work with the prox, is there any codes?

The method is already demonstrated in the manuscript. The code is not provided here, since the modification is rather simple: (1) Get the scene depth and semantics. (2) Generate multiple bodies using our model (S1 or S2), and compute their average. (3) Replace the Vposer term in the PROX fitting loss by that calculated average. Perhaps you need adjust the hyper-parameters. (4) Run the modified PROX fitting as usual.

I have tested the [PROX] (https://github.com/MohameHassan/prox) before and i find that prox fit a image cost lots of time. So I'm curious if your method can increase the running speed?

no.

Thanks a lot!

commented

Thanks for the kind reply!