RPSRNet: End-to-End Trainable Rigid Point Set Registration Network using Barnes-Hut $2^D$-Tree Representation
This work addresses the problem of efficient and accurate point cloud registration for applications like LiDAR odometry, offering a novel method that handles non-uniform densities and noise better than existing deep-learning approaches.
The paper tackles rigid point set registration by proposing RPSRNet, an end-to-end trainable deep neural network that uses a novel 2^D-tree representation and hierarchical feature embedding, achieving inference speeds of 12-15ms for point clouds up to 250K and outperforming prior methods on KITTI and ModelNet-40 datasets, with improvements such as 1.3-1.9 times better accuracy.
We propose RPSRNet - a novel end-to-end trainable deep neural network for rigid point set registration. For this task, we use a novel $2^D$-tree representation for the input point sets and a hierarchical deep feature embedding in the neural network. An iterative transformation refinement module in our network boosts the feature matching accuracy in the intermediate stages. We achieve an inference speed of 12-15ms to register a pair of input point clouds as large as 250K. Extensive evaluation on (i) KITTI LiDAR odometry and (ii) ModelNet-40 datasets shows that our method outperforms prior state-of-the-art methods - e.g., on the KITTI data set, DCP-v2 by1.3 and 1.5 times, and PointNetLK by 1.8 and 1.9 times better rotational and translational accuracy respectively. Evaluation on ModelNet40 shows that RPSRNet is more robust than other benchmark methods when the samples contain a significant amount of noise and other disturbances. RPSRNet accurately registers point clouds with non-uniform sampling densities, e.g., LiDAR data, which cannot be processed by many existing deep-learning-based registration methods.