search

zhaofuq / LOD-3DGS

LetsGo: Large-Scale Garage Modeling and Rendering via LiDAR-Assisted Gaussian(Published in SIGGRAPH Asia 2024)
MIT License
248 stars 32 forks source link

Question about large scale GS #16

Open leblond14u opened 1 month ago

leblond14u commented 1 month ago

Hi, Thanks for this amazing work !

I'm a bit new to 3DGS and I'm experimenting on large scale scenes. However whenever I'm trying to train a GS on a pseudo large scale scene with synthetic LiDAR points it seems like the GS process is making the data disapear rather than densifying it. The gaussians shrink and my GS cloud become less dense instead of densifying :/ Have you encountered the same problem in your work ?

Thanks in advance, Best regards,

Hugo

samcao0416 commented 1 month ago

Hi Hugo,

Yes we found the same case in our experiments, especially in experiments that the input LiDAR points are dense. The reason is that in some area, such as wall and ground, it is unnecessary to use as many points as point cloud for 3DGS to represent the scene.

Hope this explanation helps.

Best, Junming