The paper introduces a novel framework called Point Cloud Transformer (PCT) for point cloud learning, which is based on the Transformer architecture originally developed for natural language processing. PCT is designed to handle the irregular and unstructured nature of point clouds, making it suitable for tasks such as shape classification, part segmentation, semantic segmentation, and normal estimation. The key contributions of PCT include: 1. **Permutation Invariance**: PCT leverages the inherent permutation-invariance property of the Transformer to avoid the need for defining the order of points, which is a challenge in point cloud processing. 2. **Offset-Attention**: An optimized version of the self-attention mechanism, called offset-attention, is introduced to enhance the robustness and effectiveness of the Transformer in point cloud processing. This mechanism uses relative coordinates and a Laplacian matrix to improve the performance. 3. **Neighbor Embedding**: A neighbor embedding module is added to capture local context within the point cloud, enhancing the global feature representation. The paper also discusses the design and implementation details of PCT, including the encoder and decoder structures, and provides extensive experimental results demonstrating that PCT achieves state-of-the-art performance on various point cloud processing tasks. The experiments are conducted on datasets such as ModelNet40, ShapeNet, and S3DIS, showing superior performance compared to existing methods. The computational requirements of PCT are also analyzed, highlighting its efficiency and suitability for deployment on mobile devices.The paper introduces a novel framework called Point Cloud Transformer (PCT) for point cloud learning, which is based on the Transformer architecture originally developed for natural language processing. PCT is designed to handle the irregular and unstructured nature of point clouds, making it suitable for tasks such as shape classification, part segmentation, semantic segmentation, and normal estimation. The key contributions of PCT include: 1. **Permutation Invariance**: PCT leverages the inherent permutation-invariance property of the Transformer to avoid the need for defining the order of points, which is a challenge in point cloud processing. 2. **Offset-Attention**: An optimized version of the self-attention mechanism, called offset-attention, is introduced to enhance the robustness and effectiveness of the Transformer in point cloud processing. This mechanism uses relative coordinates and a Laplacian matrix to improve the performance. 3. **Neighbor Embedding**: A neighbor embedding module is added to capture local context within the point cloud, enhancing the global feature representation. The paper also discusses the design and implementation details of PCT, including the encoder and decoder structures, and provides extensive experimental results demonstrating that PCT achieves state-of-the-art performance on various point cloud processing tasks. The experiments are conducted on datasets such as ModelNet40, ShapeNet, and S3DIS, showing superior performance compared to existing methods. The computational requirements of PCT are also analyzed, highlighting its efficiency and suitability for deployment on mobile devices.