Abstract
Unmanned aerial vehicle (UAV)-based tracking has wide perspective applications, however, due to the limitations of computing resources, battery capacity, and maximum load of UAVs, efficiency seems more thorny and imperative as an issue than precision in UAV tracking, which explains why discriminative correlation filters (DCF)- instead of deep learning (DL)- based trackers are usually preferred in this field, as the former are known for high efficiency, whereas, the latter hardly achieve real-time tracking on a single CPU. However, without deep representation learning the precision of DCF-based trackers is extremely limited. This paper dedicates to boost the efficiency of DL-based trackers for UAV tracking by presenting a global filter pruning and proposes a self-attention module, which seeks to learn backbone features that highlight meaningful visual inter-dependencies, in order to combat the precision drop, and, importantly, to avoid the arduous process of determining layer-wise pruning ratios in the original ranked-based filter pruning method. Remarkably, self-attention is utilized here to guide the training without introducing any extra computational burden into the inference phase. Extensive experiments on four UAV benchmarks show that the proposed tracker strikes a remarkable balance between efficiency and precision and achieves state-of-the-art performance in UAV tracking.Video
Citation
@inproceedings{Liu_2022_BMVC,
author = {Mengyuan Liu and Yuelong Wang and Qiangyu Sun and Shuiwang Li},
title = {Global Filter Pruning with Self-Attention for Real-Time UAV Tracking},
booktitle = {33rd British Machine Vision Conference 2022, {BMVC} 2022, London, UK, November 21-24, 2022},
publisher = {{BMVA} Press},
year = {2022},
url = {https://bmvc2022.mpi-inf.mpg.de/0861.pdf}
}
Copyright © 2022
The British Machine Vision Conference is organised by