Point Cloud Semantic Segmentation Enhanced by Self-Training Feature Fusion

LiDAR remote sensing technology, owing to its high-precision laser pulse measurement characteristics, is capable of high-accuracy dynamic perception and modeling of 3-D spatial targets. The significant improvement in the accuracy of models based on the PointNet++ architecture has been primarily attributed to the expansion of the model size, the enhancement of the receptive fields, and the optimization of the training strategies. Inspired by the PointNeXt model, this article revisits the traditional local feature aggregation process, enhancing the receptive field of the network through an implicit decoupling approach, and jointly exploring multidimensional vector feature weighting to delve deeper into the discriminative features of the imagery. We introduce a pseudo-label generation strategy and integrate a multiclass weighted loss to achieve a comprehensive understanding of point cloud semantics and spatial information through a novel point cloud segmentation network, receptive field extension, and vector feature enhancement network (RV-Net). Compared to recent advanced fully supervised networks, RV-Net demonstrates competitive results on multiple benchmark datasets. Furthermore, experiments conducted on multiple benchmark datasets with 1% and 0.1% labeling ratios demonstrate that the proposed model outperforms various fully and weakly supervised segmentation algorithms. The results indicate that the network model proposed in this article exhibits a superior performance in the semantic segmentation of point clouds in large-scale indoor and outdoor scenarios.