SEELE: A Unified Acceleration Framework for Real-Time Gaussian Splatting

Xiaotong Huang1*, He Zhu1*, Zihan Liu12, Weikai Lin3, Xiaohong Liu1, Zhezhi He1,
Jingwen Leng12, Minyi Guo12, Yu Feng1†

1 Shanghai Jiao Tong University   2 Shanghai Qi Zhi Institute   3 University of Rochester
* Equal Contribution   † Corresponding Author
Paper Code

📹 Demo Video


🔍 Comparisons

The comparisons showcase state-of-the-art 3DGS variants (3DGS [Kerbl 2023], MiniSplatting [Fang 2024], and LightGaussian [Fan 2024])   without   and   with   SeeLe 🚀   enhancement. The FPS values are measured on an AGX Orin SoC.

3DGS
+ SeeLe 🚀
40 FPS
3DGS
[Kerbl 2023]
12FPS
3DGS
+ SeeLe 🚀
72 FPS
3DGS
[Kerbl 2023]
28 FPS
MiniSplatting
+ SeeLe 🚀
162 FPS
MiniSplatting
[Fang 2024]
92 FPS
MiniSplatting
+ SeeLe 🚀
272 FPS
MiniSplatting
[Fang 2024]
137 FPS
LightGaussian
+ SeeLe 🚀
210 FPS
LightGaussian
[Fan 2024]
75 FPS
LightGaussian
+ SeeLe 🚀
81 FPS
LightGaussian
[Fan 2024]
36 FPS
3DGS
+ SeeLe 🚀
68 FPS
3DGS
[Kerbl 2023]
25 FPS
3DGS
+ SeeLe 🚀
98 FPS
3DGS
[Kerbl 2023]
34 FPS
MiniSplatting
+ SeeLe 🚀
89 FPS
MiniSplatting
[Fang 2024]
50 FPS
MiniSplatting
+ SeeLe 🚀
176 FPS
MiniSplatting
[Fang 2024]
109 FPS
LightGaussian
+ SeeLe 🚀
76 FPS
LightGaussian
[Fan 2024]
29 FPS
LightGaussian
+ SeeLe 🚀
143 FPS
LightGaussian
[Fan 2024]
52 FPS

🔬 Sensitivity Study

(1) 📑 Cluster Number in Hybrid Preprocessing (HP, Sect. 3.3)

Cluster Number of HP

(2) 📑 Group Size in Contribution-Aware Rasterization (CR, Sect. 3.4)

Group Size in CR