DGAT-net: Dynamic Graph Attention for 3D Point Cloud Semantic Segmentation

When performing processing tasks such as semantic segmentation on point cloud data, one of the biggest challenges is how to effectively capture both the local details as well as the long-range dependencies in the data. Existing processing methods often struggle to deal with both types of information simultaneously, which leads to limited performance when processing complex scenarios. To address this problem, this paper proposes a dynamic graph attention network DGAT-net. First, a dynamic graph is constructed using a feature similarity and spatial relationship-based approach to capture the local structure and long-range semantic dependencies in point cloud data. Dynamic edge sampling is also added to update the dynamic graph to reduce computational complexity. A spatial self-attention mechanism and a graph attention mechanism are introduced to capture local detail information and global dependencies, respectively, to further enhance the model’s ability to understand complex spatial relationships. Moreover, the fusion of these two attention mechanisms combined with the gating mechanism are dynamically selected to adapt to different data characteristics and task requirements. Experimental results on public point cloud datasets show that our method achieves superior performance in semantic segmentation tasks, validating its effectiveness and practicality.