Fisher Pruning for Real-Time UAV Tracking

Despite the wide prospect of applications of unmanned aerial vehicle (UAV)-based tracking in transportation, agriculture, public security, and so on, the limitations of computing resources, battery capacity and maximum load of UAV largely hinder the deployment of deep learning (DL)-based tracking algorithms on UAV. Discriminative correlation filters (DCF)-based trackers because of their high efficiency and low resource-consuming merits have thus stood out in the UAV tracking community. However, the precision of DCF-based trackers is hardly comparable to DL-based ones in complex scenarios due to the limited representation learning ability. Filter pruning is a common technique used to deploy deep networks in edge-devices with low power and constrained resources, without compromising much on the accuracy of the model. It is probably an effective means to improve the efficiency of DL-based trackers and facilitate their deployment on UAVs. However, applying filter pruning to UAV tracking has not been well explored. A simple and effective pruning criterion is very desirable at present and may draw more attention in the UAV tracking community to model compression. In this paper, we propose to exploit Fisher pruning to compress the SiamFC++ model for UAV tracking, resulting in our proposed F-SiamFC++ tracker which demonstrates a remarkable balance between efficiency and precision. Extensive experiments on four UAV benchmarks, including UAV123@10fps, DTB70, UAVDT and Vistrone2018 (VisDrone2018-test-dev), show that the proposed F-SiamFC++ tracker achieves state-of-the-art performance.