Dynamic Memory Reconciliation for Online Action Detection

Online action detection (OAD) aims to recognize ongoing actions from streaming videos in real-time, which demands effective temporal modeling to capture both long-range dependencies and fine-grained local dynamics. The main challenge lies in the model’s inability to utilize future data, requiring it to selectively leverage the most relevant historical and current information for predictions. To address this, we propose a novel framework integrating Shuffled Global Context (SGC) module and Adaptive Local Gating (ALG) module. The SGC module dynamically reorganizes long-term contexts through cyclic shift operations, enabling efficient cross-frame interaction. Complementarily, the ALG module adaptively regulates local feature aggregation by emphasizing spatio-temporal correlations among short-term frames, which effectively suppresses irrelevant noise and highlights discriminative cues for evolving actions. Experimental results demonstrate that our method achieves competitive performance against state-of-the-art methods.