Data-centric is a novel perspective for UAV-based tracking: A new benchmark via efficient data utilization strategy

Abstract

Tracking a moving target with unmanned aerial vehicles (UAVs) poses significant challenges due to the substantial distance between the camera and target and the high relative motion. Trackers must efficiently process both appearance and motion information while adhering to the constraints of UAVs’ limited onboard computing power and real-time operational demands. Although current state-of-the-art (SOTA) UAV trackers rely on compact network structures, optimizing performance without increasing complexity remains a daunting challenge. This paper introduces a data-centric approach to enhance tracking performance in UAV environments. We first critique the limitations of existing datasets and propose a novel data mining strategy that leads to the development of the UAVSOT dataset. This dataset provides a more detailed representation for single-object tracking in UAV scenarios, effectively addressing the shortcomings of current datasets. Our experiments show that methods trained on UAVSOT significantly enhance tracking accuracy without additional computational overhead. Additionally, we compare model-centric and data-centric approaches to underscore the efficacy of our data-driven strategy in optimizing UAV trackers. The code and raw results can be found at https://github.com/caixiongyou/UAV-DC-Track.