Point Cloud Semantic Segmentation with Transformer and Multi-Scale Feature Extraction

To address the limitations of existing point cloud semantic segmentation methods in modeling long-range dependencies and adapting to multi-scale targets, we propose an improved Cylinder3D framework based on multi-scale feature extraction. Current voxel-based methods primarily rely on local 3D convolutions, limiting their effectiveness in complex scenes, while fixed-receptive-field convolutions also limit model adaptability to multi-scale targets. To address these challenges, we propose two major improvements: first, we introduce a Transformer module for global modeling in high-level feature extraction, leveraging the self-attention mechanism to enable global information exchange between different voxels, overcoming the limitations of traditional convolution's local receptive fields; second, we design a multi-scale feature extraction architecture with tri-directional dilated convolutions, namely the improved DDCM module, which combines standard convolution and dilated convolution, effectively expanding the receptive field while maintaining parameter efficiency and enhancing the processing capability for point clouds with heterogeneous density distributions. Evaluation on the SemanticKITTI dataset demonstrates that our improved model achieves an mIoU of 66.8%, a 2.4% increase over the original Cylinder3D. Our research indicates that the proposed enhancements significantly improve the accuracy of 3D point cloud semantic segmentation, though the computational complexity of the model requires further optimization, which will be addressed in future work.