vanilla gaussian splatting gradually slowing down

Anttwo / SuGaR

[CVPR 2024] Official PyTorch implementation of SuGaR: Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering

https://anttwo.github.io/sugar/

Other

1.9k stars 137 forks source link

vanilla gaussian splatting gradually slowing down #100

Open op10ds opened 5 months ago

op10ds commented 5 months ago

I have a problem when executing the following line:

python gaussian_splatting/train.py -s <path to COLMAP dataset> --iterations 7000 -m <path to the desired output directory>

the process starts with an iteration speed of about 6 it/s, then it gradually slows down to 2 it/s. In this way this first step takes more than 30 minutes, instead of few minutes as stated in the guide. I am using a laptop with rtx 3080 with 16 GB vram and 32 GB ram, i don't see any strange memory usage. I am using ubuntu on windows, through wsl2. Do you guys have any clue on what's going on here?

Anttwo commented 5 months ago

Hello @op10ds,

It's not surprising to see the optimization slowing down when running Vanilla Gaussian Spatting, as the Gaussians "densify" during optimization. In other words, the number of Gaussians progressively increases (and it sometimes increases by a lot!), so it becomes more demanding for your system.

However, I agree that 30 minutes for 7000 iterations is a lot, much more than what is expected.

Let's investigate!

May I ask the following questions:

How many points does your initial SfM point cloud have?
How many images do you have?
How does your scene look? Is it a large-scale scene?

op10ds commented 5 months ago

Hello! thanks for the feedback. Regarding your questions: 1) the file points3d.ply counts 13257 points; 2) the dataset has 30 images; 3) it is a scan of a person, so the camera moves around a subject.