search

vchoutas / smplify-x

Expressive Body Capture: 3D Hands, Face, and Body from a Single Image
https://smpl-x.is.tue.mpg.de/
Other
1.77k stars 340 forks source link

How can I reproduce the quantitative results reported in your paper? #103

Open lijiaman opened 4 years ago

lijiaman commented 4 years ago

Hi there,

Thanks for your great work. I have set up your codes and able to run it successfully. But with the default setting, I tested in your EHF dataset, there're many failure cases, I got many monster meshes...(while some are valid meshes) I noticed there're some github issues mentioning this. Could you please kindly provide guidance on how to reproduce the v2v error metrics reported in your paper? With the current setting, the number is far away from what you reported... For example, did you use ground truth camera parameters or different initialization for different images? Do you have evaluation codes for getting the v2v error, it seems generated meshes have very different root translation from ground truth mesh? Currently, the results look really bad, not that good as I expected. Hope you can help.

Thanks a lot!

Limaoxing commented 3 years ago

Hi,I'm a student of UESTC,I'm interested in 3D human body reconstruction, Can you contact me by 1835607452@qq.com,I hope we can discuss some question about 3D human body reconstruction.I'm looking forward to your reply.

Limaoxing commented 3 years ago

could you share evaluation codes for getting the v2v error,I need it badly!!!