MinkLoc3D: Point Cloud Based Large-Scale Place Recognition

The paper presents a learning-based method for computing a discriminative 3Dpoint cloud descriptor for place recognition purposes. Existing methods, suchas PointNetVLAD, are based on unordered point cloud representation. They usePointNet as the first processing step to extract local features, which arelater aggregated into a global descriptor. The PointNet architecture is notwell suited to capture local geometric structures. Thus, state-of-the-artmethods enhance vanilla PointNet architecture by adding different mechanism tocapture local contextual information, such as graph convolutional networks orusing hand-crafted features. We present an alternative approach, dubbedMinkLoc3D, to compute a discriminative 3D point cloud descriptor, based on asparse voxelized point cloud representation and sparse 3D convolutions. Theproposed method has a simple and efficient architecture. Evaluation on standardbenchmarks proves that MinkLoc3D outperforms current state-of-the-art. Our codeis publicly available on the project website:https://github.com/jac99/MinkLoc3D