Cascaded Dual-Structure Transformer Feature Fusion for Part-Based Point Cloud Segmentation

Traditional point cloud segmentation methods often over-rely on local aggregation, which struggles to effectively capture complex inter-point relationships. This limitation leads to insufficient feature extraction and limited segmentation accuracy. To address these issues, this paper proposes a Cascade Dual-structure Transformer Feature Fusion Network (CDSTNet), designed to fully leverage the detailed information and 3D representation capabilities of point clouds for precise feature segmentation of industrial part models. CDSTNet incorporates a three-layer feature extraction network to effectively capture global, local, and detailed features. These features are then deeply integrated through a self-attention fusion module, significantly enhancing the detail expression of point cloud features. Furthermore, to improve the model's understanding of point cloud topology and geometric feature extraction, a cascade dual-structure Transformer module integrating cross-attention and global attention mechanisms is constructed. On a custom dataset, CDSTNet achieves a segmentation accuracy (mIoU) of 99.71% , outperforming PointNet, PointNet++, DGCNN, and GTNet by 2.14%, 1.08%, 0.3%, and 0.1%, respectively. On a general dataset, it achieves an mIoU of 85.38%, showing improvements of 1.6%, 0.2%, and 0.1% over PointNet, PointNet++, and DGCNN. While comparable to GTNet in accuracy on the general dataset, CDSTNet doubles the training speed.