search

skhu101 / SHERF

Code for our ICCV'2023 paper "SHERF: Generalizable Human NeRF from a Single Image"
Other
302 stars 10 forks source link

Quality of RenderPeople SMPL annotation #22

Closed markkim1115 closed 11 months ago

markkim1115 commented 11 months ago

Hello, Thanks for the good work and sharing all the details, datasets.

I am checking about the RenderPeople dataset you provided.

But I see the fitted SMPL model for some subjects are inaccurate for their scale, body pose. (mostly, scale)

I attach some of the projected 3D joint visualization results

Good fitting example seq_000024-rp_amber_rigged_003

Quite good fitting example (Usually if not perfect fitting, anootations tend to have lower head position?) seq_000051-rp_beatrice_rigged_006

Scale inaccuracy example seq_000000-rp_aaron_rigged_001 seq_000381-rp_lara_rigged_003 seq_000538-rp_percy_rigged_004

Body pose inaccuracy example(Most of the annotations have good body pose fitting) seq_000130-rp_debra_rigged_005

  1. Is this my visualization code problem? or basically this dataset has inaccurate annotation?

  2. SHERF looks good for visualized samples in repo, I think does mis-aligned annotation still enough to make plausible point-level features..? But if we remove inaccurate samples or improve annotation quality, does it have rooms to improve model performance?

  3. Do you have plan to improve and provide more accurate annotations for RenderPeople?

Thanks again.

markkim1115 commented 11 months ago

Sorry, my minor visualization mistake. except the body pose fitting fail case seen in last sample i attached, i confirmed SMPL fitting is quite accurate. Closing the issue.