ODA-GS: Occlusion- and Distortion-aware Gaussian Splatting for Indoor Scene Reconstruction

In this work, we aim to address the quality degradation issues of indoor scene reconstruction using 3D Gaussian Splatting (3DGS). Existing methods enhance reconstruction quality by exploiting learned geometric priors like Signed Distance Functions (SDF), but these come with significant computational costs. We analyze the traditional 3DGS training process and identify key factors contributing to quality degradation: over-reconstruction and gradient dilution during the densification stage, and the occurrence of distorted/redundant Gaussians during the post-optimization stage. To tackle these issues, we introduce ODA-GS, a novel framework that modifies 3DGS with tailored modules. During densification, we employ occlusion-aware gradient accumulation to prevent gradient dilution and use homo-directional gradients to mitigate over-reconstruction. In the post-optimization stage, we introduce post-pruning to eliminate distorted and redundant Gaussians, thereby enhancing visual quality and reducing computational overhead. Tested on the ScanNet++ and Replica datasets, ODA-GS outperforms several baselines both qualitatively and quantitatively.