Optimization of 4D Gaussian Splatting Accelerators for Dynamic Scene Rendering
- 발행기관 서강대학교 일반대학원
- 지도교수 류성주
- 발행년도 2026
- 학위수여년월 2026. 2
- 학위명 석사
- 학과 및 전공 일반대학원 전자공학과
- 실제URI http://www.dcollection.net/handler/sogang/000000082273
- UCI I804:11029-000000082273
- 본문언어 영어
- 저작권 논문은 저작권에 의해 보호받습니다.
초록(요약문)
This paper presents Relay‑GS, an integrated algorithm and hardware framework for accelerating 4D Gaussian Splatting in dynamic scene rendering. Although 4DGS can represent motion and temporal changes effectively, most existing pipelines treat every frame independently. They repeat the same sorting and pixel computations even when adjacent frames contain nearly identical Gaussian distributions. This repetition adds unnecessary latency, reduces throughput, and limits the ability to achieve real‑time rendering. Relay‑GS reduces this overhead by adapting its processing flow to the temporal behavior of the input sequence. It reuses the depth ordering from the previous frame through Selective Gaussian Sorting, restores accuracy at fixed intervals with periodic full sorting, and lowers rasterization cost by using an incremental form of the Gaussian computation. These ideas are supported by a hardware architecture that includes a dedicated sorting unit and a parallel rasterization engine connected through a fine‑grained pipeline. Experiments with dynamic scene benchmarks show that Relay‑GS reaches up to a 1.64× speed improvement.
more목차
I. Introduction 1
II. Preliminary: Gaussian Splatting 5
2.1. 3D Gaussian Splatting (3DGS) 5
2.2. 4D Gaussian Splatting (4DGS) 8
III. Hardware Accelerator for 4DGS 10
3.1. Previous Work 10
3.2. Relay-GS 11
3.2.1. Selective Gaussian Sorting (SGS) 11
3.2.2. Periodic Selective Gaussian Sorting 15
3.2.3. Parallel Rasterization 19
3.2.4. Sorting-Rasterization Pipeline 22
3.3. Architecture 25
3.3.1. Proposed Architecture Overview 25
3.3.2. Preprocessing Unit (PPU) 25
3.3.3. Selective Gaussian Sorting Unit (SGSU) 27
3.3.4. Parallel Rasterization Unit (PRU) 28
IV. Evaluation 30
4.1. Experimental Setup 30
4.2. Rendering Quality Analysis 31
4.3. Performance Analysis 33
4.4. Energy Efficiency Analysis 34
4.5. Hardware Implementation Results 36
V. Conclusion 37
