Flow Intelligence: Robust Feature Matching Via Temporal Signature Correlation

Robust video feature matching, particularly across different modalities, remains a significant challenge as existing methods rely on spatial appearance features that are inconsistent in dynamic, real-world conditions. This paper introduces Flow Intelligence, a paradigm-shifting framework that completely abandons spatial appearance in favor of temporal motion patterns. Our training-free approach computes a binary Temporal Motion Signature (TMS) for each video block based on simple frame differencing. Correspondence is then established by efficiently correlating these binary signatures. Extensive experiments demonstrate that our method yields a denser set of matches compared to state-of-the-art learning-based approaches, while maintaining competitive regional accuracy. More importantly, it shows exceptional robustness in challenging cross-modal (e.g., RGB-Infrared) scenarios where traditional methods completely fail. By leveraging motion as the sole source of information, Flow Intelligence enables robust, efficient, and adaptable video matching without the need for large datasets or expensive training, opening new possibilities for real-world applications.