Wasserstein approximate bayesian computation for visual tracking

Abstract

In this study, we present novel visual tracking methods based on the Wasserstein approximate Bayesian computation (ABC). For visual tracking, the proposed Wasserstein ABC (WABC) method approximates the likelihood within the Wasserstein space more accurately than the conventional ABC methods by directly measuring the discrepancy between the likelihood distributions. To encode the temporal dependency among time-series likelihood distributions, we extend the WABC method to the time-series WABC (TWABC) method. Subsequently, the proposed Hilbert TWABC (HTWABC) method reduces the computational costs caused by the TWABC method while substituting the original Wasserstein distance with the Hilbert distance. Experimental results demonstrate that the proposed visual trackers outperform other state-of-the-art visual tracking methods quantitatively. Moreover, ablation studies verify the effectiveness of individual components consisting of the proposed method (e.g., the Wasserstein distance, curve matching, and Hilbert metric).