Accelerating Point Cloud Registration With Low Overlap Using Graphs and Sparse Convolutions

We present a novel correspondence-learning model for real-time registration of partially overlapping point clouds, between which the relative translation is large and hence identifying the correspondences is challenging. Our goal is to improve the feature learning for accurate correspondence establishment, which enables the promotion of registration performance significantly in terms of efficiency. This is realized by two particular designs. The first is a graph-based feature extraction module, which aggregates both inter and intra contexts of the input point clouds simultaneously for strengthening the connection between inputs. The second is a feature refinement module, which uses sparse convolutions to further widen the feature differences of dissimilar structures. The two modules reinforce each other to improve correspondence learning for robust and fast point cloud registration with low overlap. We evaluate the method on both synthetic and real-world large-scale datasets. The results in real registration tasks show that our method attains competitive registration accuracy with state-of-the-art methods, and is almost two times faster than these competing methods in some scenarios.