ABSTRACT
Three-dimensional (3D) point cloud registration is a critical topic in 3D computer vision and remote sensing. Several algorithms based on deep learning techniques have recently tried to deal with indoor partial-to-partial point cloud registration by searching the correspondences between input point clouds. However, existing correspondence-based methods are vulnerable to noise and do not adequately exploit geometric information to extract features, resulting in incorrect correspondences. In this work, we develop a novel network using correspondence confidence and overlap scores to address these challenges. Specifically, we first introduce a feature interaction module that combines spatial structure information to encode unique geometric embedding, greatly enhancing the feature perception. Furthermore, we design a corresponding point matching module, which includes a two-stage point filtering strategy. This method effectively improves the ability to identify embedded inliers from outliers and accurately remove spurious matches, thus allowing the network to focus more on the accurate correspondences of overlapping regions. Extensive experiments on different benchmark datasets indicate that our network shows superior performance of indoor point cloud registration, especially in low overlap registration, with significant improvement over state-of-the-art (SOTA) methods. Our code can be found at https://github.com/tranceok/COPRNet.
Disclosure statement
No potential conflict of interest was reported by the authors.