Discriminative object tracking by domain contrast

Abstract

Multi-domain tracking method as a novel learning paradigm improves object tracking by sharing domain information with a common backbone whilst learning private information with domain-specific layers. In that context, each individual video sequence as a specific domain serves for a domain-specific layer. In this paper, we observe an intriguing finding that target features from different domains are highly confused with each other, thus having weak discriminative ability, for lack of domain interaction. To this end, we propose a simple yet effective domain interaction training paradigm called domain contrast to boost discriminative object features by effectively using amounts of instances from all the domains in two novel aspects: (1) a light-weight memory-saving training algorithm is proposed to solve the “out-of-the-memory” problem, which paves the way to couple with previous multi-domain trackers, and (2) a composite class-balanced loss is explored to tackle a more practical imbalanced problem, which not only involves the usual class imbalance problem but also accounts for the case of the totally mere negative instances. Experiments on multiple popular tracking benchmark datasets show that our mechanism consistently achieves the tracking performance gain of both base multi-domain tracker and its real-time variant thereof, without any other changes made on the original network.