search

chenhsuanlin / bundle-adjusting-NeRF

BARF: Bundle-Adjusting Neural Radiance Fields 🤮 (ICCV 2021 oral)
MIT License
788 stars 114 forks source link

about se3_refine parameter #77

Closed pcc1016 closed 1 year ago

pcc1016 commented 1 year ago

Hello! I'm a student studying NeRF, and I have a question about the SE3_refine function in the BARF code. I'm not sure what the parameters of 6-DoF refer to. Are they related to affine transformation? I'm not entirely certain.

And regarding automatic backward, how do you know the Jacobian formula of the model and how do you compute automatic backward? I want to implement backward myself but I don't understand it well

Thank you for your research.

chenhsuanlin commented 1 year ago

Hi @pcc1016, rigid transformations (in the SE(3) space) have 6 degrees of freedom (3 for rotation and 3 for translation). The gradients are computed via autodiff with PyTorch in practice.