A Survey and Benchmark of Automatic Surface Reconstruction from Point Clouds

Area of Science:

• Computer Vision and Graphics
• Machine Learning

Background:

• Surface reconstruction from point clouds is vital but challenging due to real-world data imperfections like noise, outliers, and missing data.
• Traditional methods often rely on handcrafted priors, requiring extensive tuning, while deep learning offers data-driven approaches.

Purpose of the Study:

• To comprehensively survey and benchmark traditional and learning-based surface reconstruction methods.
• To evaluate the impact of handcrafted versus learned priors on reconstruction precision and robustness.
• To provide a standardized evaluation framework and resources for the research community.

Main Methods:

• A comparative analysis of established and recent surface reconstruction techniques.
• Standardized evaluation on diverse point cloud datasets, including those with real-world anomalies.
• Assessment of both traditional algorithms and deep learning models.

Main Results:

• Learning-based models consistently outperform traditional methods in surface quality on clean, uniform point clouds, even for novel shapes.
• Traditional methods exhibit superior robustness against noise, outliers, and missing data typical in real-world 3D acquisitions.
• The study quantifies the trade-offs between precision and robustness influenced by different prior types.

Conclusions:

• Deep learning advances surface reconstruction quality but requires further development for real-world robustness.
• Traditional methods remain valuable for their resilience in challenging acquisition scenarios.
• Open-sourced code and datasets are provided to foster future research in learning-based surface reconstruction.