Point Cloud Registration With Zero Overlap Rate and Negative Overlap Rate

Abstract

Low-overlap registration is an important subtask in point cloud registration. In this letter, we focus on two extreme states of low-overlap registration tasks: zero overlap rate and even negative overlap rate point cloud registration. Instead of filtering out non-overlapping regions, overlapping regions are created before registration to deal with the registration tasks with no overlapping regions. Specifically, a novel generative network called Regiffusion is developed on the basis of a diffusion model to predict the shape of the object after registration based on the shapes of the source and target point clouds; the complete object shape, including those of the source point cloud, is accurately predicted before registration and treated as the new target point cloud. This approach effectively creates overlapping regions between the source and target point clouds. We evaluate the developed method on the self-constructed zero-overlap dataset Pokemon-Zero, negative overlap dataset Pokemon-Neg, and the publicly available dataset ModelNet40, indoor datasets 3DMatch. Experimental results demonstrate that the presented method not only performs very welll on zero-overlap and negative overlap datasets, but also improves the registration performance on low-overlap datasets.