search

turandai / gaussian_surfels

Implementation of the SIGGRAPH 2024 conference paper "High-quality Surface Reconstruction using Gaussian Surfels".
501 stars 24 forks source link

Are the camera parameters learnable? #12

Closed alvin528 closed 4 months ago

alvin528 commented 4 months ago

Very impressive work! I'm trying to implement my ideas based on your code. But I noticed that the camera parameters q and T are nn.Parameter in https://github.com/turandai/gaussian_surfels/blob/654511a86573f311f4875c070fef8f58791a0548/scene/cameras.py#L60-L61

Are they updated with the network learning or just some unusable definitions?

turandai commented 4 months ago

Hi, we have tried to optimize the pose so the q and T are Parameter type, but they are not updating during the training as we set the default learning rate for them to 0. Although we have implemented the gradient chains from losses to the pose, this functionality has not yet been fully tested, so we suggest to turn it off. However you are welcomed to do your expirements based on it.