AbacusGS drops the per-frame depth sort in Gaussian Splatting by storing the blending order in the scene layout.
How it works: Splats are grouped into octree boxes of at most 128 splats, each stored in order along an axis chosen at write time. Each frame, only the visible boxes are sorted, and one dot product per box decides whether it is read forward or backward.
Cost: A short order-aware fine-tune converts an existing 3DGS model. Across 13 standard scenes, it loses 0.13 dB PSNR on average against an equally fine-tuned 3DGS model.
Speed: In an OpenGL viewer that sorts on the CPU, it renders frames 4.5 times faster than per-splat sorting. The paper also describes a 17-byte-per-splat format and a CUDA tile renderer that use the same order.
The code isn't out yet: the paper says it "will be released."
Sign in to comment.
TL;DR of the paper, generated automatically.
AbacusGS drops the per-frame depth sort in Gaussian Splatting by storing the blending order in the scene layout.
The code isn't out yet: the paper says it "will be released."