david-svitov / HAHA

HAHA: Highly Articulated Gaussian Human Avatars with Textured Mesh Prior
89 stars 9 forks source link

custom data process #5

Closed R132 closed 1 month ago

R132 commented 2 months ago

Hi! I want to try your work on my own data. You mentioned "We use the data from InstantAvatar and finetuned their pre-processed poses to get SMPL-X fits." for SnapshotPeople. Did you use VIBE to estimate the SMPLX parameters? According to your paper.

david-svitov commented 2 months ago

Hi! To obtain SMPL-X parameters we used SMPLify-x: https://github.com/vchoutas/smplify-x From our experience, optimization-based methods provide higher accuracy than feed-forward methods.

Recommendations not related to the article: You can use VIBE to initialize SMPL-X parameters and then fine-tune them using SMPLify-X. To obtain more accurate SMPL-X fits, you can also modify SMPLify-X by adding a segmentation loss as suggested in this article: https://samsunglabs.github.io/MoRF-project-page/