Visual tracking by dynamic matching-classification network switching. (November 2020)

• Record Type: Journal Article
• Authors: Li, Peixia; Chen, Boyu; Wang, Dong; Lu, Huchuan
• Is Part Of: Pattern recognition. Volume 107(2020:Nov.)
• Publication Date: 2020-11
• Subjects: Visual Tracking -- Deep Learning -- Ensemble learning; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2020.107419
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: A dynamic matching-classification switching framework is proposed to effectively integrate the matching, classification and verification networks. A meta classifier is learned to adapt quickly to the current appearance changes through one iteration one training sample, which speeds up the online tracking. Extensive experiments on two popular benchmarks show that the proposed tracker achieves good performance compared with recent state-of-the-art methods. Abstract: Existing deep trackers can be roughly divided into either matching-based or classification-based methods. The formers are fast but not very robust; while the latter ones introduce more discriminative information but often very slow. In this work, we present a novel real-time robust tracking method to take full use of the benefits from both kinds of networks. First, we propose a matching-classification network switching (MCS) framework to integrate the matching, classification, verification networks and conduct dynamic switching among them. Second, to speed up online update, we devlop a meta learning method as a critical component in our classification network. The meta classifier is trained offline to obtain general discriminative ability and updated online to the current frame just through one iteration. Extensive experiments are conducted on two popular benchmark datasets. Both qualitative and quantitative evaluations show that our tracker performs favorably against other state-of-the-art trackers with real-time performance.