The paper introduces a new Cross-Modal Information-Guided Network (CMIGNet) for point cloud registration. Traditional methods rely solely on features from points, leading to limitations such as lack of texture information and inadequate global feature perception. The proposed CMIGNet uses cross-modal information for better global shape perception. It incorporates projected images from point clouds and fuses the cross-modal features using an attention mechanism. Two contrastive learning strategies are employed, focusing on overlapping features and correspondences between 2D and 3D features. A mask prediction module is also proposed to identify keypoints in the point clouds. The network shows superior registration performance in experiments on several benchmark datasets.
Publication date: 3 Nov 2023
Project Page: https://github.com/IvanXie416/CMIGNet
Paper: https://arxiv.org/pdf/2311.01202
