Token Context-Aware Reference Token Selection for Visual Object Tracking
A visual object-tracking pipeline that improves the use of reference-frame context by selecting only the most informative tokens instead of treating every image patch equally. A Token Context Memory module autoregressively accumulates high-quality spatio-temporal target tokens while discarding redundant background tokens, and a Unidirectional Token Attention mechanism links the retained reference tokens to the search frame for robust cross-frame association and target localization.
2501.00758
In visual object tracking, contextual information from multiple reference frames helps perceive the target state, but naive frame-context methods treat every patch equally and are hurt by noisy, redu…