GSReuse: Temporally Adaptive Screen-Space Reuse for Accelerating 3D Gaussian Splatting.

Area of Science:

• Computer Graphics
• Real-time Rendering
• Virtual Reality

Background:

• 3D Gaussian Splatting (3DGS) enables high-fidelity novel view synthesis.
• Independent frame rendering in 3DGS video sequences causes computational redundancy, impacting VR performance.
• Existing frame reuse methods are not suitable for point-based 3DGS rendering.

Purpose of the Study:

• To develop a lightweight accelerator for 3DGS rendering that reuses computations across consecutive frames.
• To improve the runtime efficiency of 3DGS without retraining scene representations.
• To facilitate the practical deployment of 3DGS in demanding applications like VR.

Main Methods:

• GSReuse operates in screen space, requiring minimal modifications to existing 3DGS pipelines.
• It estimates motion vectors and warps content from previous frames.
• A tile-based filtering and masking strategy identifies reusable regions, enabling skipped rendering operations.

Main Results:

• GSReuse significantly accelerates 3DGS rendering while maintaining high visual fidelity.
• It outperforms state-of-the-art video frame reuse methods in both image quality and latency.
• The method demonstrates effectiveness across multiple benchmark datasets.

Conclusions:

• GSReuse offers a practical solution to the runtime inefficiency of 3DGS video rendering.
• Its lightweight and drop-in nature makes it suitable for real-time applications, especially VR.
• The proposed method enhances the feasibility of real-time 3DGS applications by reducing computational costs.