Point Cloud Fitting Based on High-Accuracy Segmentation

Point cloud fitting is a classic problem in computational geometry, this paper proposes a deep learning method to fit implicit functions for segmented point cloud without requiring point cloud patch type information.Since the fitting error is highly related to the accuracy of point cloud segmentation, making it crucial to build an accurate model for segmenting point cloud into different primitive patches.The proposed method introduces PCBNET: Proximal Points and Curvature Based Primitive Segmentation Network, achieving stateof-the-art results of SegMiou (Segmentation Mean Intersection over Union).The network introduces proximal points as labels; these are points adjacent to others that have different primitive labels.Moreover, the estimated Gaussian and mean curvatures are involved to enhance the prediction accuracy of the proximal points' labels.Experiments show that using the encoded features of predicted proximal points enhances the model's segmentation performance.After obtaining initial segmentation labels using mean-shift clustering, authors employ DBSCAN (Density-Based Spatial Clustering of Applications with Noise) to further improve segmentation accuracy.Finally, authors fit the segmented point cloud using a new deep implicit functions learning method, which was originally designed for B-Rep solid fitting. CCS Concepts Computing methodologies Point-based models.