search

River-Zhang / SIFU

[CVPR 2024 Highlight] Official repository for paper "SIFU: Side-view Conditioned Implicit Function for Real-world Usable Clothed Human Reconstruction"
https://river-zhang.github.io/SIFU-projectpage/
MIT License
196 stars 8 forks source link

Thuman2.0 evaluation is coming out differently #39

Open jingi0614 opened 4 weeks ago

jingi0614 commented 4 weeks ago

Thank you for your great contribution to this paper.

I discovered that some data have a scale difference between the SMPLX normal map and the Thuman scan files during data preprocessing. This seems to be affecting the evaluation. How should I resolve this? 120_inter Most of the data are correctly scaled, but there are some instances where the scales do not match. I would appreciate any assistance you can provide. @River-Zhang

jingi0614 commented 3 weeks ago

The issue was resolved after reprocessing the dataset using the new SMPL-X from THuman2.0. However, there is still a slight discrepancy in performance compared to the results reported in the SiFu paper. Could you help identify what might be causing this difference?

When visualizing the 4-directional feature maps used in SiFu, I've noticed a difference between the feature maps generated with the provided pretrained weights and those generated with the weights from the model I trained. I'm not sure why this discrepancy occurs, and I would greatly appreciate your assistance in understanding the possible reasons. @River-Zhang

merged_image

sifu

This is the performance of the model I trained. Thuman2.0 chamfer: tensor(0.7175) p2s: tensor(0.7261) NC: tensor(0.0449) {'cape-easy-NC': 0.0338255874812603, 'cape-easy-chamfer': 0.7092598080635071, 'cape-easy-execution_time': 0.4802453056971232, 'cape-easy-p2s': 0.6766804456710815, 'cape-hard-NC': 0.0384397991001606, 'cape-hard-chamfer': 0.8673397302627563, 'cape-hard-execution_time': 0.4780437970161438, 'cape-hard-p2s': 0.7746955752372742}