Low-overlap Point Matching Network based on Offset Gragh Attention

Yifan Ge, Xianhui Liu

Abstract: Rigid point cloud registration has become an essential task in robotics and computer vision. The main challenges are the extraction of key points and the correspondences, especially on the low-overlap point clouds. This paper proposes an iterative algorithm for correspondences based on graph attention features (LoPMN) to address the situation. To this end, our approach uses row-column alternating normalization and an overlapping region prediction module to obtain the soft assignments of point correspondences. The results of the experiments on the 3DMatch dataset show that LoPMN performs better than recent deep-learning methods.

External IDs:doi:10.1145/3679409.3679436