Optimal visual tracking using Wasserstein transport proposals

Abstract

We propose a novel visual tracking method based on the Wasserstein transport proposal (WTP). In this study, we theoretically derive the optimal proposal function in Markov chain Monte Carlo (MCMC) based visual tracking frameworks. For this objective, we adopt the optimal transport theory in the Wasserstein space and present a new transport map that can transform from a simple proposal distribution to the optimal target distribution. To find the best transport map, we conduct an additional Monte Carlo simulation. Experimental results demonstrate that the proposed method outperforms other state-of-the-art visual tracking methods. The proposed WTP can be substituted with conventional proposal functions in an MCMC framework, and thus can be plugged into any existing MCMC-based visual tracker.