An Adversarial Framework for Multi-View Action Recognition with Cross-View Generalization

Multi-view action recognition remains a challenging task due to significant appearance and motion discrepancies across camera viewpoints. In this paper, we propose a novel framework that integrates Mamba-based temporal modeling with adversarial learning to improve cross-view generalization. Specifically, our method extracts motion representations from multiple views using a dual-stream structure and introduces static perturbations during training to suppress reliance on view-specific biases. A consistency constraint is further applied to align the semantic representations across views. Extensive experiments on multiple multi-view datasets demonstrate that our approach significantly outperforms existing state-of-the-art methods under cross-view evaluation, highlighting its effectiveness and generalization capability.