EgoNN: Egocentric Neural Network for Point Cloud Based 6DoF Relocalization at the City Scale

The letter presents a deep neural network-based method for global and local descriptors extraction from point cloud acquired by rotating 3D LiDAR. can be used two-stage 6DoF relocalization. First, course position is retrieved finding candidates with the closest descriptor in database of geo-tagged clouds. Then, pose between query estimated matching using robust estimator such as RANSAC. Our has simple, fully convolutional architecture based on sparse voxelized representation. It efficiently extract set keypoints large clouds tens thousand points. code pretrained models are publicly available project website.