A Large Scale Benchmark of Person Re-Identification

Abstract

Unmanned aerial vehicles (UAVs)-based Person Re-Identification (ReID) is a novel field. Person ReID is the task of identifying individuals across different frames or views, often in surveillance or security contexts. At the same time, UAVs enhance person ReID through their mobility, real-time monitoring, and ability to access challenging areas despite privacy, legal, and technical challenges.To facilitate the advancement and adaptation of existing person ReID approach to the UAV scenarios, this paper introduces a baseline along with two datasets, i.e., LSMS and LSMS-UAV. Both datasets have the following key features: (1) LSMS: Raw videos captured by a network of 29 cameras deployed across complex outdoor environments. LSMS-UAV: captured by 1 UAV. (2) LSMS: Videos span both winter and spring seasons, encompassing diverse weather conditions and various lighting conditions throughout different times of the day. (3) LSMS: Including the largest number of annotated identities, comprising 7730 identities and 286,695 bounding boxes. LSMS-UAV: comprising 500 identities and 2000 bounding boxes. Comprehensive experiments demonstrate LSMS’s excellent capability in addressing the domain gap issue when facing complex and unknown environments. The LSMS-UAV dataset verifies that UAV data has strong transferability to traditional camera-based data.