Temporal action detection in long, untrimmed videos often stumbles on imprecise boundaries and sparse annotations, producing low-quality proposals. In a new arXiv preprint, Rui Chen and colleagues propose HyrCap, a query-based framework that treats dense predictions as an over-complete space of temporal event hypotheses. The method's name points to its two core ideas: hybrid feature fusion and rank calibration of proposal confidence.

The hybrid component combines complementary video features that carry different localization cues. Rank-calibration then adjusts proposal ranking confidence so that it reflects localization reliability and competition among candidates. Three mechanisms support this: Cross-Representation Consensus Calibration, Localization Quality Estimation, and Proposal Redundancy Suppression.

The authors also extend HyrCap to semantic action understanding, feeding calibrated proposals into LLM-based event analysis. They define a staged evaluation protocol that measures event localization, segment-level semantic understanding, and the combined task separately. On public benchmarks, they report state-of-the-art performance across multiple metrics.