search

hbb1 / 2d-gaussian-splatting

[SIGGRAPH'24] 2D Gaussian Splatting for Geometrically Accurate Radiance Fields
https://surfsplatting.github.io
Other
2.11k stars 144 forks source link

Slower training compared to 3DGS #176

Open Golbstein opened 6 days ago

Golbstein commented 6 days ago

Hi I believe the slower training time is mainly due to excessive scene densification, my assumption is that because 2DGS have less degrees of freedom we have to apply more densification to properly represent the scene, and that leads to almost x2 slower training compared to original 3DGS repo.

I made sure I'm using the same params in both trainings..

3DGS. (27 iterations/sec) Training progress: 45%|██████████▌ | 9000/20000 [05:00<06:47, 27.00it/s, Loss=0.0302305, Depth Loss=0.0000000, Points=494349]

2DGS (14iter/sec) and more points Training progress: 45%|███████▌ | 9000/20000 [08:32<12:46, 14.35it/s, Loss=0.03731, distort=0.00000, normal=0.00000, Points=591914]

However, the rasterizer is indeed faster

hbb1 commented 3 days ago

@Golbstein Hi, which 3DGS version did you use? I notice that the new 3DGS repo have integrated some advances that significant improve speed, such as fused SSIM, ADAM Optimizer, and per-splat backward. In the future, these features may be also integrated into this repo if needed.

Golbstein commented 2 days ago

Yes I do use the most updated repos