SOTA: Sequential Optimal Transport Approximation for Visual Tracking in Wild Scenario

In this study, we introduce a probabilistic visual tracking method tailored for wild scenarios, where tracking environments experience abrupt changes over time. In probabilistic visual tracking, particularly when utilizing sequential Monte Carlo (MC) sampling, the careful choice of a proposal function is critical for attaining precise tracking results, where optimal transport techniques can assist in reducing the variance from important sampling and inducing a suitable proposal function. However, if a tracked object undergoes significant changes in appearance over time, it creates a large discrepancy between the proposal function and its target distribution. This situation presents the topological and representational challenges for conventional optimal transport techniques. To address this problem, the proposed visual tracker leverages the benefits of both MC sampling and optimal transport and presents the Sequential Optimal Transport Approximation (SOTA) visual tracker. For this, our method involves transforming the proposal function into its target distribution across successive temperature steps through sequential MC sampling. These MC steps can reduce the topological and representational burden on the optimal transport. Within each successive temperature step, one distribution is projected onto another one via optimal transport. By using optimal transport, the method can mitigate the variance from importance sampling. The experimental results demonstrate that the proposed method considerably outperforms other state-of-the-art methods across several benchmark dataset, particularly in tracking environments where abrupt changes occur.