Multimodal-guided prototype calibration and temporal coherence-aware hybrid matching for few-shot action recognition