Fast Point Cloud Registration using Semantic Segmentation

Deep learning has recently delivered relatively high quality semantic segmentation of visual and point-cloud data. This paper is primarily concerned with the use of such semantic segmentation for point cloud registration. In particular, we are motivated by the need to speed up, for large scale data sets, algorithms for registration that guarantee optimality (in terms of maximising consensus). That semantic information can help prune bad hypotheses for point matches is rather obvious, and we demonstrate one such relatively simple approach by modifying a recent optimal registration algorithm [6] to take advantage of semantic information. However, we also make another contribution in proposing a novel variation of deep learning approaches to point cloud registration. Again, our motivation is handling large data sets and in this case we are able to provide an algorithm that achieves on par with state-of-the-art performance on the semantic segmentation task. In short, we have shown how to speed up both the generation of the semantic information, and how to use that semantic information to speed up point cloud registration, in the context of large scale point cloud data-sets.