GTLR-GS: Geometry-Texture Aware LiDAR-Regularized 3D Gaussian Splatting for Realistic Scene Reconstruction

Abstract

Recent advances in 3D Gaussian Splatting (3DGS) have enabled real-time, photorealistic scene reconstruction. However, conventional 3DGS frameworks typically rely on sparse point clouds derived from Structure-from-Motion (SfM), which inherently suffer from scale ambiguity, limited geometric consistency, and strong view dependency due to the lack of geometric priors. In this work, a LiDAR-centric 3D Gaussian Splatting framework is proposed that explicitly incorporates metric geometric priors into the entire Gaussian optimization process. Instead of treating LiDAR data as a passive initialization source, 3DGS optimization is reformulated as a geometry-conditioned allocation and refinement problem under a fixed representational budget. Specifically, this work introduces (i) a geometry-texture-aware allocation strategy that selectively assigns Gaussian primitives to regions with high structural or appearance complexity, (ii) a curvature-adaptive refinement mechanism that dynamically guides Gaussian splitting toward geometrically complex areas during training, and (iii) a confidence-aware metric depth regularization that anchors the reconstructed geometry to absolute scale using LiDAR measurements while maintaining optimization stability. Extensive experiments on the ScanNet++ dataset and a custom real-world dataset validate the proposed approach. The results demonstrate state-of-the-art performance in metric-scale reconstruction with high geometric fidelity.

Figures (8)

• Figure 1: Our custom-designed backpack-mounted mobile system enables the synchronized capture of images and LiDAR point clouds for reconstruction in textureless and geometrically degenerate real-world scenes. (a) Our method on data captured with the custom MLS device. (b) 3DGS results on SfM-based data.
• Figure 2: GTLR-GS Overview. i) The input data consists of dense LiDAR point clouds and RGB images. ii) Geometric and texture complexity is computed to guide point cloud sampling for 3DGS initialization. iii) Dynamically adjusts the splitting intensity based on local surface complexity. iv) Depth maps are extracted from the registered point cloud map and used as regularization constraints to supervise 3DGS training. v) The trained 3DGS results include the Gaussian sphere distribution structure and novel view synthesis performance.
• Figure 3: (a) Raw scanned point cloud from the ScanNet++ dataset. (b) Randomly downsampled point cloud. (c) Geometry-texture-aware point cloud allocation under a fixed budget.
• Figure 4: Curvature-adaptive splitting enables finer-grained refinement in geometrically complex regions, thereby preserving detailed structural geometry.
• Figure 5: Initialized with randomly sampled LiDAR point clouds. (a) Point cloud distribution of the original 3DGS. (b) Point cloud distribution after applying the geometry-aware splitting strategy.