DeTarNet: Decoupling Translation and Rotation by Siamese Network for Point Cloud Registration

Point cloud registration is a fundamental step for many tasks. In this paper, we propose neural network named DetarNet to decouple the translation t and rotation R, so as overcome performance degradation due their mutual interference in point registration. First, Siamese Network based Progressive Coherent Feature Drift (PCFD) module proposed align source target points high-dimensional feature space, accurately recover from alignment process. Then Consensus Encoding Unit (CEU) construct more distinguishable features set of putative correspondences. After that, Spatial Channel Attention (SCA) block adopted build classification finding good Finally, obtained by Singular Value Decomposition (SVD). way, decouples estimation rotation, resulting better both them. Experimental results demonstrate that improves on indoor outdoor scenes. Our code will be available https://github.com/ZhiChen902/DetarNet.