SparseSurf: Sparse-View 3D Gaussian Splatting for Surface Reconstruction

TL;DR

SparseSurf tackles sparse-view surface reconstruction by augmenting 3D Gaussian Splatting with stereo-derived geometric priors and multi-view feature consistency. The Stereo Geometry-Texture Alignment provides metric depth and normal supervision from rendered stereo pairs, while Pseudo-Feature Enhanced Geometry Consistency enforces multi-view coherence through feature distillation and pseudo-view supervision. The approach yields state-of-the-art performance on DTU, BlendedMVS, and Mip-NeRF360 under sparse-view constraints, improving both surface accuracy and rendering quality. This work advances practical 3D surface reconstruction with sparse inputs by integrating stereo geometry, texture alignment, and learned feature regularization into Gaussian-based representations, with implications for efficient 3D modeling in real-world capture scenarios.

Abstract

Recent advances in optimizing Gaussian Splatting for scene geometry have enabled efficient reconstruction of detailed surfaces from images. However, when input views are sparse, such optimization is prone to overfitting, leading to suboptimal reconstruction quality. Existing approaches address this challenge by employing flattened Gaussian primitives to better fit surface geometry, combined with depth regularization to alleviate geometric ambiguities under limited viewpoints. Nevertheless, the increased anisotropy inherent in flattened Gaussians exacerbates overfitting in sparse-view scenarios, hindering accurate surface fitting and degrading novel view synthesis performance. In this paper, we propose \net{}, a method that reconstructs more accurate and detailed surfaces while preserving high-quality novel view rendering. Our key insight is to introduce Stereo Geometry-Texture Alignment, which bridges rendering quality and geometry estimation, thereby jointly enhancing both surface reconstruction and view synthesis. In addition, we present a Pseudo-Feature Enhanced Geometry Consistency that enforces multi-view geometric consistency by incorporating both training and unseen views, effectively mitigating overfitting caused by sparse supervision. Extensive experiments on the DTU, BlendedMVS, and Mip-NeRF360 datasets demonstrate that our method achieves the state-of-the-art performance.

Figures (11)

• Figure 1: Comparison of sparse-view novel-view synthesis and surface reconstruction on DTU and Mip-NeRF360. Our SparseSurf achieves the best performance on both surface reconstruction and rendering in sparse-view setting.
• Figure 2: The framework of SparseSurf. (a) Stereo Geometry-Texture Alignment. We estimate and update stereo-view images to generate binocular priors for geometry supervision. (b) Pseudo-Feature Enhanced Geometry Consistency. To mitigate overfitting and enhance multi-view consistency, we introduce Pseudo-view Feature Consistency and Train-view Feature Alignment.
• Figure 3: Qualitative comparison of reconstruction results on the DTU with little-overlap and large-overlap sparse setting.
• Figure 4: Qualitative rendering comparison on DTU with sparse-view NVS setting.
• Figure 5: Comparison on BlendedMVS.