Robust Gaussian Splatting via Depth Reliability Masking and Edge-Aware Optimization

Abstract 3D Gaussian Splatting (3DGS) achieves real-time, high-fidelity rendering through explicit Gaussian primitives and efficient rasterization techniques. However, the absence of geometric information often leads to artifacts around object edges and weakly textured regions. Although existing methods attempt to optimize geometric representation by imposing depth constraints, their efficacy remains limited due to interference from pervasive sensor noise. To address this, we propose a novel optimization framework integrating edge-aware mechanisms with depth reliability detection. Specifically, our approach employs multi-scale local depth statistics and gradient information to strategically exclude depth loss computations in ambiguous background edge regions. Simultaneously, it utilizes neighborhood depth consistency to construct a robust reliability mask that actively suppresses the influence of depth outliers. Experiments on the TUM-RGBD dataset demonstrate that our method significantly mitigates blurring and visual artifacts while improving the evaluation metrics holistically.