Du Juan ; Wang Rui ; Cremers Daniel

Abstract

For relocalization in large-scale point clouds, we propose the first approachthat unifies global place recognition and local 6DoF pose refinement. To thisend, we design a Siamese network that jointly learns 3D local feature detectionand description directly from raw 3D points. It integrates FlexConv andSqueeze-and-Excitation (SE) to assure that the learned local descriptorcaptures multi-level geometric information and channel-wise relations. Fordetecting 3D keypoints we predict the discriminativeness of the localdescriptors in an unsupervised manner. We generate the global descriptor bydirectly aggregating the learned local descriptors with an effective attentionmechanism. In this way, local and global 3D descriptors are inferred in onesingle forward pass. Experiments on various benchmarks demonstrate that ourmethod achieves competitive results for both global point cloud retrieval andlocal point cloud registration in comparison to state-of-the-art approaches. Tovalidate the generalizability and robustness of our 3D keypoints, wedemonstrate that our method also performs favorably without fine-tuning on theregistration of point clouds that were generated by a visual SLAM system. Codeand related materials are available athttps://vision.in.tum.de/research/vslam/dh3d.