Sparse-View Surface Reconstruction using Gaussian Splatting through High-Confidence Depth Propagation with Normal Priors
Proposes a Gaussian Splatting-based scene optimization with normal-guided depth propagation, outperforming SOTA in sparse-view 3D surface reconstruction.
Key Findings
Methodology
This paper introduces a scene-specific 3D Gaussian Splatting framework integrating normal-guided depth propagation and abnormal depth edge-aware regularization. It estimates depth confidence via multi-view geometric and photometric consistency, then propagates high-confidence depths guided by surface normals to low-confidence regions. An edge-aware loss suppresses depth discontinuities caused by Gaussian discreteness. The overall loss combines depth propagation, edge regularization, normal prior, and rendering errors, enabling high-fidelity surface reconstruction from sparse views. The approach effectively addresses depth ambiguity, especially in textureless or occluded areas, and accelerates convergence compared to prior methods.
Key Results
- On DTU and Tanks-and-Temples datasets, the method reduces Chamfer distance to 1.02 and improves F1 scores to 0.293 under sparse views, outperforming FatesGS and MAtCha. It maintains robustness across 5, 10, and 20 views, producing smoother, more complete surfaces.
- Quantitative evaluations show significant improvements in geometric accuracy, with ablation studies confirming the critical role of depth propagation and edge regularization. Visual results demonstrate finer details and fewer artifacts than competing methods.
- The method achieves a balance of high reconstruction quality and computational efficiency, with training times comparable to existing 3DGS-based approaches, making it practical for real-world applications.
Significance
This work advances sparse-view 3D surface reconstruction by effectively extending depth information and suppressing artifacts caused by Gaussian discretization. It bridges the gap between dense-view methods and real-world scenarios where limited images are available, enabling applications in cultural heritage, robotics, and industrial inspection. The integration of normal priors and depth confidence estimation provides a robust framework that can adapt to various scene complexities, marking a significant step toward practical, high-fidelity 3D modeling from minimal input data.
Technical Contribution
The paper introduces a novel scene-specific optimization framework combining multi-view geometric and photometric cues to estimate depth confidence. It employs normal-guided depth propagation to extend reliable depth cues into uncertain regions, coupled with an abnormal depth edge-aware regularization to ensure surface smoothness. This approach significantly improves the stability and accuracy of sparse-view surface reconstruction, surpassing existing methods in both quality and efficiency. The framework's modular design facilitates integration with other implicit or explicit representations, broadening its applicability.
Novelty
This is the first work to incorporate normal-guided depth propagation within a 3D Gaussian Splatting framework for sparse-view surface reconstruction. It uniquely combines multi-view consistency with normal priors to extend depth cues beyond high-confidence regions, addressing depth ambiguity and discontinuities caused by Gaussian discreteness. The edge-aware regularization further distinguishes this approach, making it more robust to scene complexity and view sparsity compared to prior methods that rely solely on direct depth supervision or regularization.
Limitations
- The local planar assumption underlying depth propagation may fail in highly non-planar or complex geometries, limiting accuracy in such scenarios.
- Dependence on accurate normal estimation means that errors in normal prediction can propagate into depth supervision, affecting overall reconstruction quality.
- The method's performance diminishes in scenes with severe occlusion or extreme sparsity, requiring further integration of global constraints or additional modalities for improvement.
Future Work
Future research will explore integrating global geometric priors and multi-modal data, such as texture and point clouds, to enhance robustness in complex scenes. Efforts will also focus on real-time implementation and scalability for dynamic environments, broadening the method's application in robotics, AR/VR, and autonomous navigation. Additionally, improving normal estimation accuracy and extending the framework to handle non-Lambertian surfaces remain promising directions.
AI Executive Summary
Reconstructing detailed 3D surfaces from sparse views remains a core challenge in computer vision, especially when limited input images hinder depth and geometric cues. Existing methods often require dense views or extensive pre-training, which are impractical for many real-world applications such as cultural heritage preservation or robotic navigation. This paper introduces a novel scene-specific approach based on 3D Gaussian Splatting, which leverages normal-guided depth propagation and abnormal depth edge-aware regularization to overcome these limitations.
The key innovation lies in estimating a depth confidence map through multi-view geometric and photometric consistency, then propagating high-confidence depths guided by surface normals into uncertain regions. This process effectively extends reliable depth cues, enabling the model to infer accurate surface geometry even in textureless or occluded areas. To address depth discontinuities caused by the discrete nature of Gaussians, an edge-aware smoothing loss is incorporated, ensuring smooth, coherent surfaces.
Extensive experiments on DTU and Tanks-and-Temples datasets demonstrate that this method outperforms state-of-the-art techniques like FatesGS and MAtCha, achieving lower Chamfer distances and higher F1 scores under sparse-view conditions. Visual comparisons reveal finer details and more complete surfaces, validating the effectiveness of the depth propagation and regularization strategies.
This approach significantly advances the field by enabling high-fidelity surface reconstruction from minimal input data, reducing computational costs, and improving robustness. Its potential applications span industrial inspection, cultural heritage digitization, and autonomous systems, where rapid, accurate 3D modeling is essential. Despite current limitations in highly complex scenes and dependence on normal accuracy, future work aims to incorporate global constraints and multi-modal data to further enhance performance. Overall, this research paves the way for practical, efficient sparse-view 3D reconstruction in diverse real-world scenarios.
Deep Dive
Glossary
3D Gaussian Splatting (3DGS)
一种基于高斯分布的点云渲染技术，用于高效生成和优化三维场景模型。
本文采用3DGS作为场景优化的基础框架。
深度置信图 (Depth Confidence Map)
通过多视几何和光度一致性估算的深度可靠性指标，用于引导深度传播。
用以识别高置信区域，指导深度扩展。
法线先验 (Normal Prior)
预训练模型预测的场景表面法线信息，用于引导深度传播和正则化。
作为深度扩展的引导信息。
异常深度边缘 (Abnormal Depth Edge)
深度不连续或不合理的边缘区域，需通过正则正则化抑制。
通过边缘感知正则进行平滑处理。
多视几何一致性 (Multi-view Geometric Consistency)
不同视角下深度和几何信息的一致性，用于估算深度置信。
作为深度置信估算的重要依据。
Open Questions Unanswered questions from this research
- 1 如何在极端遮挡或复杂几何场景中进一步提升深度传播的准确性仍未充分解决，需结合全局几何和多模态信息以增强鲁棒性。
- 2 深度传播假设局限于局部平面，难以应对非平面或大范围复杂结构，未来需引入更复杂的模型以突破此限制。
Applications
Immediate Applications
工业检测
利用该方法快速重建机械零件或结构的三维模型，实现缺陷检测和质量控制。
文化遗产保护
用少量照片还原古建筑或文物的细节，便于数字存档和修复。
Long-term Vision
自主机器人导航
在复杂环境中实现实时高精度三维重建，支持自主避障和路径规划。
Abstract
3D reconstruction from sparse views is a challenging task in 3D computer vision. Recent studies on 3D Gaussian Splatting (3DGS) have achieved remarkable results with sparse views in novel view synthesis, yet reconstructing high-quality geometric surfaces from sparse views remains a challenge, due to the limited geometry clues and the discreteness of Gaussians. In this paper, we propose a novel 3DGS-based method for high-fidelity surface reconstruction from sparse views. Our key insight is to introduce a normal-guided depth propagation approach, which can extend depth information from high-confidence regions to constrain the depth in low-confidence areas. Additionally, we propose an abnormal depth edge-aware regularization to address depth discontinuities caused by the discreteness of Gaussians. Extensive experiments on DTU and Tanks-and-Temples datasets demonstrate that our method outperforms the state-of-the-art methods in sparse view surface reconstruction. Project page: https://hanl2010.github.io/DP-GS.