GCMTN: Low-Overlap Point Cloud Registration Network Combining Dense Graph Convolution and Multilevel Interactive Transformer

A single receptive field limits the expression of multilevel features in point cloud registration, leading to pseudo-matching objects with similar geometric structures low-overlap scenes, which causes a significant degradation registration performance. To handle this problem, network that incorporates dense graph convolution and mutilevel interaction Transformer (GCMTN) pursuit better performance scenes is proposed paper. In GCMTN, feature aggregation module designed for expanding points fusing at multiple scales. make pointwise more discriminative, combining Multihead Offset Attention Cross refine internal perform interaction. filter out undesirable effects outliers, an overlap prediction containing factor matching also determining match ability predicting region. The final rigid transformation parameters are generated based on distribution GCMTN was extensively verified publicly available ModelNet ModelLoNet, 3DMatch 3DLoMatch, odometryKITTI datasets compared recent methods. experimental results demonstrate significantly improves capability extraction achieves competitive scenes. Meanwhile, has value potential application practical remote sensing tasks.