search

taconite / IntrinsicAvatar

[CVPR 2024] IntrinsicAvatar: Physically Based Inverse Rendering of Dynamic Humans from Monocular Videos via Explicit Ray Tracing
https://neuralbodies.github.io/IntrinsicAvatar/
MIT License
87 stars 5 forks source link

Wired training results on People_snapshot dataset #9

Open SherlockSunset opened 1 month ago

SherlockSunset commented 1 month ago

I just trained on male-3-casual, but the result seems wired as follows:

企业微信截图_17283907341789

  1. seems there are some noisy region near the foot and hand
  2. the albedo quality is not so good May I know is this normal or any way can I improve this?
SherlockSunset commented 1 month ago

Besides I also tried female-3-sport, the results seems also not so good. case_2

taconite commented 1 month ago

I just tried a clean installation myself and could not produce your issue. Did you make sure that you followed the installation steps in both README.md and DATASET.md?

A potential cause for this kind of artifacts is that the input body pose estimation is noisy. Please make sure you have used the refined body pose estimation provided by this repo instead of the poses provided by the original PS dataset.

SherlockSunset commented 1 month ago

I just tried a clean installation myself and could not produce your issue. Did you make sure that you followed the installation steps in both README.md and DATASET.md?

A potential cause for this kind of artifacts is that the input body pose estimation is noisy. Please make sure you have used the refined body pose estimation provided by this repo instead of the poses provided by the original PS dataset.

Hi, are u referring to refine the poses of original PS dataset using the file scripts/preprocess_CAPE.py?

taconite commented 1 month ago

I just tried a clean installation myself and could not produce your issue. Did you make sure that you followed the installation steps in both README.md and DATASET.md? A potential cause for this kind of artifacts is that the input body pose estimation is noisy. Please make sure you have used the refined body pose estimation provided by this repo instead of the poses provided by the original PS dataset.

Hi, are u referring to refine the poses of original PS dataset using the file scripts/preprocess_CAPE.py?

The refined poses are already in the load folder. If you follow the PeopleSnapshot preprocessing instructions in DATASET.md you should get the proper data for PeopleSnapshot.

SherlockSunset commented 1 month ago

Hi Shaofeng, thank you for your prompt response. Actually I’ve mostly followed the instructions in DATASET.md, but I’ll double-check my conda environment as well as the processed PeopleSnapshot dataset just to be sure.