Surface Reconstruction From Point Clouds: A Survey and a Benchmark.

Area of Science:

• Computer Vision
• Computer Graphics
• Geometric Modeling

Background:

• Reconstructing continuous 2D surfaces from discrete point clouds is a challenging, ill-posed problem.
• Practical depth scanning introduces imperfections like noise, outliers, and missing data.
• Existing reviews lack comprehensive benchmarking on common datasets.

Purpose of the Study:

• To review and benchmark existing surface reconstruction methods, particularly in the context of deep learning.
• To introduce a large-scale dataset for evaluating surface reconstruction techniques.
• To assess method robustness against common scanning imperfections and generalization capabilities.

Main Methods:

• Development of a comprehensive benchmarking dataset with synthetic and real-scanned data.
• Empirical evaluation of various surface reconstruction methods on the benchmark dataset.
• Analysis of method performance concerning robustness to imperfections and generalization to complex shapes.

Main Results:

• Classical surface reconstruction methods demonstrate superior robustness and generalization compared to some deep learning approaches.
• Deep learning methods are increasingly popular but do not universally outperform established techniques.
• Significant challenges persist in handling point set misalignment, missing data, and outliers.

Conclusions:

• The developed benchmark dataset provides valuable insights for practitioners and researchers.
• Classical methods offer strong performance, challenging the dominance of newer deep learning techniques.
• Further research is needed to address fundamental issues in point cloud surface reconstruction.