CCAG: End-to-End Point Cloud Registration

Abstract

Point cloud registration is a crucial task in computer vision and 3D reconstruction, aiming to align multiple point clouds to achieve globally consistent geometric structures. However, traditional point cloud registration methods face challenges when dealing with low overlap and large-scale point cloud data. To overcome these issues, we propose an end-to-end point cloud registration method called CCAG. The CCAG algorithm leverages the Cross-Convolution Attention module, which combines cross-attention mechanism and depth-wise separable convolution to capture relationships between point clouds and integrate features. Through cross-attention computation, this module establishes associations between point clouds and utilizes depth-wise separable convolution operations to extract local features and spatial relationships. Furthermore, the CCAG algorithm introduces Adaptive Graph Convolution MLP, which dynamically adjusts node representations based on the positions of nodes in the graph structure and features of neighboring nodes, enhancing the expressive power of nodes through MLP. Our algorithm demonstrates competitive performance in multiple benchmark tests, including 3DMatch/3DLoMatch, KITTI, ModelNet/ModelLoNet, and MVP-RG.