search

fudan-zvg / 4d-gaussian-splatting

[ICLR 2024] Real-time Photorealistic Dynamic Scene Representation and Rendering with 4D Gaussian Splatting
MIT License
608 stars 44 forks source link

Some prolems with feature_rest when training #50

Closed tingzhongyue closed 3 weeks ago

tingzhongyue commented 2 months ago

Thank you for your great work and contribution.

When I tried to train 4dgs, I found that although active_sh_degree_t = 2, sh[16:] which should calculate 4D spherindrical harmonics was still 0.

I trained it on sear_steak dataset, and no hyperparameters were modified.

Do you ever meet some problems like this, and do you know any method to solve it?

Alexander0Yang commented 2 months ago

Hi @tingzhongyue! Thank you for pointing this out. You can turn on convert_SHs_python to reproduce the experiments in our paper. Now the CUDA version does not exactly match the behavior of the python version, although they don't significantly affect performance. I will fix them as soon as possible.