3D Point Cloud Classification Segmentation Model Based on Improved PointNet++

In response to the problem that PointNet++often ignores some smaller features during the process of max pooling, resulting in the loss of important features, this paper introduces a method combining the average pooling layer and the improved X-Conv operator on the basis of the existing network model PointNet++to replace the max pooling layer and solve the problem of feature loss. Regarding the improvement of X-Conv, Manhattan metric is used instead of Euclidean metric to solve the problem of dimensionality explosion during model operation. A three-layer X-Conv structure is set up to better extract local features, and finally, average pooling is combined with X-Conv to better focus on global features. The experimental results show that the overall accuracy (OA) and average accuracy (mAcc) of our algorithm on the ModelNet40 dataset reached 93.4% and 90.6%, respectively, and the average IoU on the ShapeNet dataset reached 86.2%. Our algorithm effectively improves the accuracy of point cloud classification and segmentation.