Approximating LiDAR Point Clouds with Gaussian Surfels

Montag, 7. Juli 2025

Evaluating Gaussian Splatting using LiDAR Point Clouds:
This project explores a hybrid 3D reconstruction pipeline that blends image-based techniques with dense LiDAR data to produce more accurate and complete models.
This project, first use Structure-from-Motion to generate a sparse point cloud from photographs, then integrate high-resolution LiDAR scans for precise geometry.
Gaussian splatting primitives are initialized with this fused dataset and refined via differentiable rendering to improve shape and appearance.
It benchmark this method against purely image-based approaches using PSNR, SSIM, and perceptual metrics.
Results show LiDAR-enhanced initialization yields sharper edges, fewer holes, and better overall fidelity—ideal for VR and robotics.