Abstract: Aiming at the problem that the segmentation accuracy of point cloud based on PointNet++ are not high at geometric plane level, the sampling algorithm, feature extraction and feature transfer process of the existing segmentation framework were studied. The farthest point sampling algorithm based on curvature was used to obtain more sampling points at the junction of part surface and surface. Combining the PointSIFT and PointCNN segmentation frameworks, the X-transform matrix was used to make the point cloud features spatially invariant. Neighbors in different directions were selected for combination by SIFT grouping. The SIFT-X convolution operator was designed to allow the network to encode point cloud features in different directions, so the network representation ability was improved. The results indicate that the above methods can improve the accuracy and mIoU value of point cloud surface element segmentation.
Key words: deep learning; point cloud segmentation; spatial transformation invariance