Dual attention focus network for few-shot skeleton-based action recognition

Few-shot action recognition is a challenging yet practically significant problem that involves developing a model capable of learning discriminative features from a small number of labeled samples to recognize new action categories. Current methods typically infer spatial relationships either within or across skeletons to learn action representations, but this often results in features with insufficient discriminability and ineffective attention to critical body parts. To address these limitations, we propose DAF-Net, a novel framework that employs focal attention to jointly model intra-skeleton and inter-skeleton relationships, enhancing discriminative feature learning in few-shot skeleton-based action recognition. Unlike traditional methods that focus solely on intra-skeleton dependencies or inter-skeleton structures, DAF-Net dynamically integrates both components via focal attention, enhancing key body part representation and refining features, particularly in data-scarce conditions. Furthermore, DAF-Net incorporates an enhanced prototype generation strategy, optimizing class prototype formation via cosine similarity weighting to further improve feature discriminability in multi-shot scenarios. In temporal matching, cosine similarity evaluates local feature similarity within skeleton sequences, capturing directional variations of specific joints over time. Extensive experiments on three benchmark datasets (NTU-T, NTU-S, and Kinetics-skeleton) confirm significant performance gains, validating the effectiveness of DAF-Net.