Deep Point Cloud Edge Reconstruction via Surface Patch Segmentation.

Area of Science:

• Computer Graphics
• Geometric Modeling
• Machine Learning

Background:

• Parametric edge reconstruction from point clouds is crucial for 3D modeling.
• Current methods struggle with sparse and inaccurately sampled edge points, leading to fitting errors.
• Existing deep learning approaches often overlook non-edge areas, limiting reconstruction completeness.

Purpose of the Study:

• To develop a robust framework for precise and complete parametric edge reconstruction from point cloud data.
• To overcome limitations of sparse edge points by leveraging contextual information from segmented surface patches.
• To enable accurate detection of corners and reconstruction of topologically connected 3D wireframe models.

Main Methods:

• Proposed a novel two-stage framework utilizing surface patch segmentation.
• Developed PCER-Net (Point Cloud Edge Reconstruction Network) for simultaneous surface patch segmentation, edge point detection, and normal prediction.
• Implemented a joint optimization module for reconstructing complete 3D wireframes using network outputs and geometric optimization.

Main Results:

• Segmented patches enabled accurate parametric edge fitting, even with non-uniformly distributed points.
• Corners were naturally detected from segmented patches, contributing to complete wireframe reconstruction.
• The proposed method demonstrated superior performance in reconstructing precise and complete 3D wireframes compared to existing techniques.

Conclusions:

• Leveraging neighboring surface patches significantly enhances parametric edge reconstruction accuracy and completeness.
• The proposed two-stage framework effectively addresses the challenges posed by sparse and noisy point cloud data.
• The developed method provides a robust solution for generating high-quality 3D wireframe models, supported by a new versatile dataset.