Xia Yan ; Xu Yusheng ; Li Shuang ; Wang Rui ; Du Juan ; Cremers Daniel ; Stilla Uwe

Abstract

We tackle the problem of place recognition from point cloud data andintroduce a self-attention and orientation encoding network (SOE-Net) thatfully explores the relationship between points and incorporates long-rangecontext into point-wise local descriptors. Local information of each point fromeight orientations is captured in a PointOE module, whereas long-range featuredependencies among local descriptors are captured with a self-attention unit.Moreover, we propose a novel loss function called Hard Positive Hard Negativequadruplet loss (HPHN quadruplet), that achieves better performance than thecommonly used metric learning loss. Experiments on various benchmark datasetsdemonstrate superior performance of the proposed network over the currentstate-of-the-art approaches. Our code is released publicly athttps://github.com/Yan-Xia/SOE-Net.