Multi-local-task learning with global regularization for object tracking

Shunli Zhang, Yao Sui, Sicong Zhao, Xin Yu, Li Zhang

Abstract: Highlights • A multi-local-task learning with global regularization method is proposed. • Tracking is formulated as a novel multi-local-task learning problem. • The closed-form solution to the multi-local-task learning is derived. • The local tasks can assist in addressing the occlusion problem and guide the update. • The global regularization can constrain the relationship of the local tasks. Abstract In this paper, we propose a novel multi-local-task learning with global regularization (GR-MLTL) method for object tracking. In our formulation, the tracking task is decomposed into several local tasks by dividing the whole target into several fragments, and the final tracking result is obtained by combining the local tasks. Specifically, we propose a global regularization term and inject it into the objective function of the multi-local-task learning formulation, and derive a closed-form solution. In our method, both the local and the global properties are embedded into a unified framework, which can not only retain the integral structure of the target by the global regularization, but also address the occlusions effectively by the local tasks. Experimental results demonstrate that our method is robust and achieves comparable performance to many state-of-the-art methods.