DSTE-Net: Dual-Scale Spatial-Temporal Excitation Network for Dynamic Gesture Recognition

Abstract

Dynamic gesture recognition is a crucial technology for achieving natural human-computer interaction and holds broad application prospects in fields such as virtual reality, smart home, and digital entertainment. However, existing methods often lack a unified mechanism to simultaneously capture subtle local motion features and global semantic context, which are both essential for accurate recognition. To address this limitation, we propose a Dual-scale Spatial-Temporal Excitation (DSTE) module that explicitly integrates temporal and spatial features across both local and global scales, enhancing the network to effectively capture fine-grained motion and long-range dependencies. The plug-and-play DSTE module can be seamlessly incorporated into standard 2D architectures such as ResNet-50, forming a more expressive framework referred to as DSTE-Net, while adding only 1.99 GFLOPs. Extensive experiments on the EgoGesture, IPNHand, and LD-ConGR datasets using only RGB videos as input demonstrate that DSTE-Net achieves superior performance compared with state-of-the-art methods.