search

facebookresearch / localrf

An algorithm for reconstructing the radiance field of a large-scale scene from a single casually captured video.
MIT License
976 stars 60 forks source link

why detach viewdirs ? #24

Closed tb2-sy closed 1 year ago

tb2-sy commented 1 year ago

Nice work! When I read the code model/TensorBase.py, I found that line 619 is somewhat confused. Why do you want to detach viewdirs? valid_rgbs = self.renderModule( xyz_sampled[app_mask], viewdirs[app_mask].clone().detach(), app_features, refine #why viewdirs detach? )

ameuleman commented 1 year ago

Hi, We wanted the pose estimation to get optimized from the difference in sampling location and not from non Lambertian effects. Our intuition is that gradients for pose estimation should come from the geometry and not from the appearance. The impact of propagating view directions' gradients in the render module has not been formally evaluated in the context of our work.