TCAN: Text-Guided Category-Aware Network for Online Action Detection

Online Action Detection (OAD) requires recognizing ongoing actions under strict causality, where only past and current frames are accessible. Existing methods rely heavily on visual features and often suffer from inter-class ambiguity due to limited semantic guidance. Directly incorporating vision–language models is challenging because of the modality gap between streaming visual features and static text embeddings. To address this, we propose a Text-guided Category-Aware Network (TCAN) that integrates category-level semantics into frame representations. TCAN introduces a Category-Aware Enhancer with bidirectional cross-attention to enable mutual refinement between visual features and text embeddings. We further design a cross-modal alignment loss to enhance category discrimination. Extensive experiments show that TCAN consistently improves performance across multiple OAD benchmarks, demonstrating its effectiveness in reducing inter-class ambiguity.