Geometrically aware transformer for point cloud analysis

摘要

With the increasing use of 3D point cloud data in autonomous driving, robotic perception, and remote sensing, efficient and accurate point cloud analysis remains a critical challenge. This study presents PointGA, a lightweight Transformer-based model that enhances geometric perception for improved feature extraction and representation. First, PointGA expands the original 3D coordinates into various geometric information, introducing more prior knowledge into the network. Second, a trigonometric position encoding suitable for point clouds is designed, which effectively enhances the expressive capability of positional information and performs preliminary feature extraction through pooling layers, significantly improving the model's robustness across various tasks. Finally, a positional differential self-attention (PDA) mechanism with linear complexity is developed to optimize feature representation and achieve efficient computation. Experimental results demonstrate that PointGA achieves 87.6% overall accuracy on the ScanObjectNN dataset for classification and 66.2% mean intersection over union(mIoU) on the S3DIS Area 5 dataset for segmentation, outperforming existing methods. These findings highlight the model's capability to balance efficiency and accuracy, offering a promising solution for point cloud analysis tasks.