search

UCDvision / compact3d

Official implementation of "Compact3D".
https://ucdvision.github.io/compact3d/
MIT License
146 stars 7 forks source link

FPS Confused #13

Closed 18128381510 closed 1 month ago

18128381510 commented 1 month ago

Great job. But I am a bit confused. compactGS can reduce memory, but how does it improve rendering speed (FPS)? compactGS represents n Gaussians with m cluster centers (m < n), but during rendering, m cluster centers replicate n Gaussians for rendering, so the number of Gaussians during rendering remains n. Doesn't this mean that the computational workload remains unchanged? Therefore, I am quite puzzled as to why the paper mentions an improvement in rendering speed ?

klnavaneet commented 1 month ago

Hi @18128381510. You are right, just using vector quantization will not have much impact on rendering speed. The reason for increase in FPS is the reduction in the number of Gaussians in a scene due to opacity regularization. The regularization forces a lot of Gaussians to be highly transparent which are then pruned without impacting rendering quality. Since the number of Gaussians reduces by 2-3x, we see a similar boost in speed.

18128381510 commented 1 month ago

Thanks