ROS-GS: Relightable Outdoor Scenes With Gaussian Splatting

Area of Science:

• Computer Vision
• Computer Graphics
• Photogrammetry

Background:

• Outdoor scene decomposition is challenging due to unbounded scenes and variable lighting.
• Existing methods like Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS) face limitations in computational overhead and lighting representation accuracy.

Purpose of the Study:

• To propose ROSGS, an efficient two-stage pipeline for reconstructing relightable outdoor scenes.
• To overcome the limitations of existing methods in rendering efficiency and relighting accuracy.

Main Methods:

• Leveraging monocular normal priors for initial geometry reconstruction using 2D Gaussian Splatting (2DGS).
• Implementing a hybrid lighting model with spherical Gaussians for sunlight and Spherical Harmonic coefficients for skylight.
• Decomposing scene texture and lighting based on the reconstructed geometry.

Main Results:

• ROSGS achieves state-of-the-art performance in relighting outdoor scenes.
• Demonstrates superior relighting accuracy and rendering efficiency compared to existing methods.
• Provides an efficient and accurate geometric foundation for scene reconstruction.

Conclusions:

• ROSGS offers an effective solution for efficient and accurate relighting of outdoor scenes.
• The hybrid lighting model successfully captures complex outdoor lighting conditions.
• The proposed pipeline significantly advances the capabilities of Gaussian Splatting for scene reconstruction and relighting.