Gradient-Driven Point Allocation for Depth-Guided 3D Gaussian Splatting in Indoor Scenes

Abstract

High-precision 3D environmental perception is crucial for autonomous systems to navigate and interact effectively within complex indoor environments. However, achieving high-granularity 3D reconstruction and high-fidelity rendering remains challenging. Although 3D Gaussian Splatting (3DGS) enables efficient photorealistic rendering, its geometric inconsistency limits its ability to reconstruct high-accuracy scenes. Standard 3DGS relies on sparse Structure-from-Motion (SfM) initialization, which often leads to uneven point distributions and floaters, particularly in weak-texture areas like plain walls. Furthermore, existing methods based on depth regularization or geometric constraints usually fail to address the fundamental sparsity of the initialization. To address these issues, we propose a lightweight gradient-driven point allocation strategy for depth-guided Gaussian splatting that bypasses traditional SfM. By utilizing RGB-D priors, we obtain a high-quality, structurally continuous initial Gaussian distribution through multi-view depth back-projection and texture-aware density control. Our method incorporates a bidirectional re-projection check to eliminate transient noise and occlusions, ensuring geometric consistency. We introduce a dual-grid allocation scheme that adaptively distributes Gaussians based on spatial frequency, mandating density in high-gradient regions while enforcing sparsity in flat areas. Quantitative and qualitative experiments on the ScanNet++ dataset demonstrate that our method achieves superior reconstruction accuracy and rendering quality. Notably, our approach achieves state-of-the-art performance with a significantly reduced Gaussian count, decreasing from 1.70M to 0.39M in representative scenes, thereby mitigating unnecessary point cloud bloat and reducing memory overhead without sacrificing structural integrity.