Learning improved representations in encoder-decoder networks for point cloud registration via point interaction modules

Point cloud registration focuses on exploring and exploiting feature similarities to match multiple point clouds of a scene from different viewpoints. However, in many real-world scenarios, points are missing (e.g., multiple LiDAR scans), or there is a substantial variation of points in different viewpoints (e.g., from 3D reconstruction), making registration more difficult. In this paper, we argue that the quality of point clouds should be improved to ease the matching process. Using point cloud completion techniques, we aim to enhance the learned low-dimensional point representations in state-of-the-art encoder-decoder architectures. We propose two modules with easy drop-in integration to support the registration: a point moving and point attention offset module. The point moving module refines the positions of irregular point clouds to strengthen the alignment, whereas the point attention offset module improves the likelihood of point matches. Both modules promote interactions among the implicit point representations to improve matching accuracy. To demonstrate the validity of our idea, we modified those representations in two popular encoder-decoder networks. In our evaluation, we use the datasets 3DMatch, 3DLoMatch, and ModelNet40. When incorporating our proposed modules, we achieve 90.6% RR, 72.4% IR with 5000 samples on 3DMatch and 69.3% RR, 43.5% IR with 5000 samples on 3DLoMatch as a new state-of-the-art. Our findings suggest that improving the point cloud quality via learned point representation benefits point cloud registration.