Indoor Point Cloud Object Segmentation Based on Direction Coding and Hole Sampling

To address the problem that existing methods ignore directional information when processing local features and cannot effectively extract point cloud neighborhood features due to the limitation of convolutional kernel size, this paper proposes a point cloud segmentation method. Firstly, by combining directional encoding and hole convolution, the local receptive field of the network is maximized. secondly, graph convolutional neural networks are utilized to mine information within local neighborhoods of points. subsequently, a neighborhood feature extraction layer is employed to automatically weight and fuse neighborhood features into a more representative single feature point. finally, in conjunction with a spatial attention mechanism, the connections between distant points are enhanced. The results of object segmentation experiments on the S3DIS dataset show that the OA and mIoU of the proposed method are, respectively, 1.3 percentage points and 4.0 percentage points higher than PointWeb, 0.6 percentage points and 0.7 percentage points higher than the baseline method RandLA-Net, and the use of hole sampling and directional coding can effectively improve the semantic segmentation accuracy of point clouds.