DBDNet:Partial-to-partial point cloud registration with dual branches decoupling

Abstract

Point cloud registration plays a crucial role in various computer vision tasks. In this paper, we concentrate on two aspects of the point cloud registration problem: rotation–translation decoupling and partial overlapping. To eliminate the error introduced by rotation and translation mutual interference, we propose a dual branches structure that produces separate correspondence matrices for rotation and translation. These dual branches are guided by distinct loss functions, facilitating independent calculation of rotation and translation. For partial-to-partial registration, we consider overlap prediction as a preordering task before the registration procedure. We propose an overlap predictor to initially identify common parts within the source and target point clouds. Subsequently, only these overlapping points are routed to the registration module. To accurately predict pointwise masks, we employ an overlap predictor that benefits from explicit feature interaction introduced by the powerful attention mechanism. Additionally, we design a multi-resolution feature extraction network to capture patterns in various scales, thereby enabling our model to exploit both local and global features. Experimental results on both synthetic and real-world datasets validate the effectiveness of our proposed method. The source code is available at https://github.com/Shi-Qi-Li/DBDNet.