search

graphdeco-inria / gaussian-splatting

Original reference implementation of "3D Gaussian Splatting for Real-Time Radiance Field Rendering"
https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/
Other
14.64k stars 1.92k forks source link

BUG ->Training --iteration 30000 ++ #1058

Open jaco001 opened 2 days ago

jaco001 commented 2 days ago

I want to refine models for more then 30000 steps. At start all is fine. But at the end I got always same ply model as 30000 saved to eg 40k folder.

With different datasets are the same effect. 40k, 50k, 60k no matter, splat count is the same as in written 30k file.

I tried combination with --test_iterations --save_iterations --checkpoint_iterations No luck. Any Idea? Hard count in the code or I miss some parameters?

AsherJingkongChen commented 2 days ago

It is supposed to be, not a bug.

It is because the densification process, which changes the count of splats, stops at 15000 iterations by default. You can increase the parameter: https://github.com/graphdeco-inria/gaussian-splatting/blob/54c035f7834b564019656c3e3fcc3646292f727d/arguments/__init__.py#L94