EgoMamba: Fusing hand-guided spatial cues and multi-scale temporal context for efficient egocentric action recognition

Abstract

Egocentric action recognition requires modeling hand-object interaction regions and their temporal evolution while maintaining high computational efficiency. Existing methods are mainly based on CNNs or Transformers. CNN-based methods are limited in long-range dependency modeling, whereas Transformer-based methods often incur high computational and memory costs. Recent state space models such as Mamba offer an efficient alternative for sequence modeling, but directly applying them to video understanding remains insufficient for capturing action-critical spatial interactions and non-strictly causal temporal context in egocentric scenes. To address these limitations, we propose EgoMamba, an ef..