Multiview Transformer for Multiway Point Cloud Registration

Abstract

To accurately reconstruct complete 3D scenes from multiple point cloud inputs, estimating reliable correspondences and rigid transformation parameters between these point clouds is essential. Deep learning methods have recently seen significant advancements in point cloud registration, mainly through leveraging attention-based transformer architectures. However, these approaches are primarily designed for pairwise registration, which presents additional challenges in scenarios for reconstructing a scene from multiple inputs. In this thesis, we introduce a Multiview Transformer designed with an emphasis on memory efficiency. Our approach is developed through a novel loss function, particularly well-suited for multiway input point cloud scenarios. We have enhanced the performance by effectively updating the coarse-level voxel features for robust registration. This advancement allows our method to surpass the performance of existing state-of-the-art techniques.