DDRNet: Fast point cloud registration network for large-scale scenes

Efficient registration for large-scale scene point clouds is a basic capability that is essential for many real-world intelligence applications, such as autonomous driving or simultaneous localization and mapping. Because a large-scale scene point cloud has various characteristics such as large volumes of data, irregular local density, and limited overlapping, most existing approaches can only be trained and operated using small-scale point clouds. In this work, we propose a deep direct registration network, named DDRNet, to efficiently align the point clouds of large-scale scenes. The DDRNet consists of three parts: a local-spatially aware encoder that can efficiently aggregate posture information containing both local and spatial features; an attentional weighting module that allows our network to self-adaptively focus on overlapping areas; and a pyramid transformation decoder used to estimate transformation based on features having different resolutions. Also, we propose a partially-subsampled strategy, which is compatible with any learning-based registration method, to enable the network to be trained and tested in a self-supervised manner. We comprehensively validated our network's efficiency and robustness using four datasets: the ModelNet40, 3Dmatch, S3DIS, and the KITTI odometry datasets. The results demonstrate that our approach is more efficient and performs better than state-of-the-art methods, including both classical and learning-based methods, on large-scale scene data and object point cloud data, but with higher robustness to the variations in point density and overlapping. The efficiency and low registration error will make DDRNet attractive for substantial applications relying on a point cloud registration task.