From source to target, point cloud registration solves for a rigid body transformation that aligns the two point clouds. IterativeClosest Point (ICP) and other traditional algorithms require a long registration time and are prone to fall into local optima. Learning-based algorithms such as Deep ClosestPoint (DCP) perform better than those traditional algorithms and escape from local optimality. However, they are still not perfectly robust and rely on the complex model design due to the extracted local features are susceptible to noise. In this study, we propose a lightweight point cloud registration algorithm, DeepMatch. DeepMatch extracts a point feature for each point, which is a spatial structure composed of each point itself, the center point of the point cloud, and the farthest point of each point. Because of the superiority of this per-point feature, the computing resources and time required by DeepMatch to complete the training are less than one-tenth of other learning-based algorithms with similar performance. In addition, experiments show that our algorithm achieves state-of-the-art (SOTA) performance on both clean, with Gaussian noise and unseen category datasets. Among them, on the unseen categories, compared to the previous best learning-based point cloud registration algorithms, the registration error of DeepMatch is reduced by two orders of magnitude, achieving the same performance as on the categories seen in training, which proves DeepMatch is generalizable in point cloud registration tasks. Finally, only our DeepMatch completes 100% recall on all three test sets.
Keywords: 3D vision; algorithms; datasets; point cloud registration; transformation.
Copyright © 2022 Qi, Wu, Ge and Sun.