hku-mars/GS-SDF

What it solves

GS-SDF addresses the challenge of achieving both photorealistic rendering and high-granularity surface reconstruction in 3D scene mapping. It solves the problem of geometric inconsistency often found in standard Gaussian Splatting, by integrating LiDAR data to ensure the reconstructed surfaces are physically grounded and geometrically consistent.

How it works

The system combines Gaussian Splatting with a neural Signed Distance Field (SDF). It uses accurate LiDAR point clouds to train a neural SDF, which provides a manifold geometry field. This SDF then guides the initialization of Gaussian primitives (SDF-based Gaussian initialization) and provides geometric regularization during the training process to ensure the resulting rendering and reconstruction are consistent with the physical geometry of the scene.

Who it’s for

This project is for researchers and developers working in 3D reconstruction, SLAM (Simultaneous Localization and Mapping), and robotic perception, specifically those using LiDAR-visual systems.

Highlights

• LiDAR-Visual Synergy: Combines the strengths of LiDAR for geometry and Gaussian Splatting for photorealistic appearance.
• SDF-Guided Initialization: Uses a neural SDF to place Gaussian primitives in physically grounded locations.
• Headless Rendering: Supports batch fly-through rendering along arbitrary camera paths for evaluation and comparison.
• Multi-camera Support: Compatible with multi-camera datasets and Colmap-format data.