Two-stage modality-graphs regularized manifold ranking for RGB-T tracking

This paper proposes a two-stage modality-graphs regularized manifold ranking algorithm to learn a robust object representation for RGB-Thermal tracking. The bounding box of the tracked object is represented with a set of patches, which are described by RGB-thermal features. We assign each patch with a weight to specify its importance in describing the object, and also each modality with a weight to reflect modal reliability. These weights are jointly optimized via manifold ranking on the modality-graphs with patches as nodes. Moreover, we develop a two-stage ranking strategy to mitigate the effects of inaccurate patch weights initialization. The object representation is then updated by imposing the modality weights and the patch weights on the extracted patch features, and the object location is finally predicted by adopting the structured SVM. Extensive experiments on the standard benchmark dataset GTOT suggest that the proposed tracker outperforms several state-of-the-art RGB-T tracking methods.