2.3.1. PointNet++ Segmentation Model

PointNet++ [36] is a deep neural network capable of directly processing disordered point cloud data. It is an upgraded version of PointNet [37] that addresses the limitations of the PointNet network with regard to local feature extraction. The network is primarily used for point cloud classification and segmentation. In this experiment, the segmentation network of PointNet++ is utilized for branch–leaf segmentation.

The segmentation network is comprised of an encoder and a decoder. The encoder is primarily responsible for the point cloud downsampling process, and extracting the local features of the point cloud by setting up multiple Set Abstraction structures. The Set Abstraction is composed of sampling, grouping, and PointNet modules, which eventually output a point cloud with global features. The decoder, on the other hand, is responsible for the upsampling process. The downsampled points are restored through distance-based interpolation, and the characteristics of each point are calculated based on the KNN, which are then sent to Softmax to achieve point-level classification. Figure 2 shows the structure of the segmentation network.

Structure of PointNet++ segmentation network. N represents the number of points, K represents the number of groups, d represents the coordinate dimension, and C represents the feature dimension.