search

ajhamdi / ges-splatting

Original reference implementation of "GES : Generalized Exponential Splatting for Efficient Radiance Field Rendering" [CVPR 2024]
https://abdullahamdi.com/ges/
Other
198 stars 7 forks source link

GES & GS rasterization code difference & debug #8

Closed wuyingxiong closed 2 months ago

wuyingxiong commented 2 months ago

Dear author！ Thanks for your work! I have some questions about the differences in the code between these two models. 1.Are the cuda parts of these two models, namely diff-gaussian-rasterization and simple-knn, the same? 2.GES has an additional _shape parameter compared to GS, mainly by changing the size of the scales input into the rasterization. If I want the β value of the GEFs in GES to be as large as possible, which hyperparameter should I adjust?

ajhamdi commented 2 months ago

1- the diff-gaussian-rasterization is modified from the original code to allow for the new shape parameter $\beta$. However the new modified rasterizer still allows running the original Gaussian by fixing $\beta=2$ 2- $\beta$ is learned and can not be set to a specific value , if you need to fix it to a specific value you need to remove the gradients and change its initialization to your desired high value.

reachomk commented 1 month ago

Hi, where are these changes made? I ran a diffchecker on forward.cu and backward.cu from submodules/diff-gaussian-rasterization in both ges-splatting and the standard repo, and I cannot seem to find any differences.