Geometric Enhancement in 3D Gaussian Splatting for Sparse-view Scene Reconstruction

Qi Li, Zheng WANG, Xing Xu, Ning Xie, Jingkuan Song, Heng Tao Shen

Keywords: Geometric Enhancement, 3DGS, Sparse-view Scene Reconstruction

Abstract: Although recent sparse-view scene reconstruction with 3D Gaussian Splatting (3DGS) like InstantSplat has made significant progress, it still suffers from geometric inconsistencies, including floating artifacts, incomplete surface reconstruction, and unstable Gaussian primitives, which significantly degrade both visual quality and geometric fidelity. Additionally, the inaccurate camera pose will also exacerbate these issues. Therefore, we present a novel geometric enhancement framework for 3DGS including geometric regularization and multi-view consistency enforcement to fundamentally address these limitations. Specifically, our approach is composed of three key components: Side-view Inconsistency Filtering (SIF) at initialization, Local Depth Regularization (LDR), and Anisotropy-aware Shape Regularization (ASR) at training. The SIF module mainly leverages multi-view information to eliminate geometrically inconsistent points, which aims to reduce floating artifacts and improve surface coherence. LDR enforces spatial consistency by identifying and penalizing regions with high geometric uncertainty through patch-based depth correlation analysis. By controlling the opacity and scale ratio, ASR can constrain Gaussian primitives to geometrically plausible shapes, preventing degenerate elongated structures. Extensive experiments on two widely used datasets demonstrate the effectiveness and superiority of our geometric enhancement when compared to pose-free methods and even pose-known baselines.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 7213