Correspondence-Free Point Cloud Registration Via Feature Interaction and Dual Branch [Application Notes]

• Published in: IEEE computational intelligence magazine Vol. 18; no. 4; pp. 66 - 79
• Main Authors: Wu, Yue, Liu, Jiaming, Yuan, Yongzhe, Hu, Xidao, Fan, Xiaolong, Tu, Kunkun, Gong, Maoguo, Miao, Qiguang, Ma, Wenping
• Format: Magazine Article
• Language: English
• Published: Washington IEEE 01-11-2023; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Adaptive systems; Correlation; Data models; Dual branch; Feature extraction; feature interaction; Image reconstruction; Measurement; Optimization; Outliers (statistics); Point cloud compression; reconstruct point cloud; Registration; Resistance; Robustness; self-adaptive representation; Solution space; Tutorials
• ISSN: 1556-603X; 1556-6048; 1556-6048; 1556-603X
• DOI: 10.1109/MCI.2023.3304144

Point cloud registration, which effectively coincides the source and target point clouds, is generally implemented by geometric metrics or feature metrics. In terms of resistance to noise and outliers, feature-metric registration has less error than the traditional point-to-point corresponding geometric metric, and point cloud reconstruction can generate and reveal more potential information during the recovery process, which can further optimize the registration process. In this paper, CFNet, a correspondence-free point cloud registration framework based on feature metrics and reconstruction metrics, is proposed to learn adaptive representations, with an emphasis on optimizing the network. Considering the correlations among the paired point clouds in the registration, a feature interaction module that can perceive and strengthen the information association between point clouds in multiple stages is proposed. To clarify the fact that rotation and translation are essentially uncorrelated, they are considered different solution spaces, and the interactive features are divided into two parts to produce a dual branch regression. In addition, CFNet with its comprehensive objectives estimates the transformation matrix between two input point clouds by minimizing multiple loss metrics. The extensive experiments conducted on both synthetic and real-world datasets show that our method outperforms the existing registration methods.