Point Cloud Semantic Segmentation Network Based on Adaptive Convolution and Attention Mechanism

Abstract

To solve the problem of negligible spatial distribution of point cloud data in the existing deep learning-based 3D points cloud semantic segmentation model, a network architecture based on adaptive convolution and attention mechanisms for semantic segmentation is proposed. Firstly, a set of feature matrices are constructed to form a convolution kernel, which can represent geometric topological structure. Secondly, the spatial correlation in local neighborhood is modeled to flexibly simulate the spatial structure of point cloud and extract effective fine-grained local features. Finally, the attention mechanism is introduced to aggregate local neighborhood features, which can enhance the spatial expression of strong correlation neighborhoods. The proposed method is tested on the S3DIS dataset, and the experimental results show that the proposed method has the best performance among the similar methods.