An Attention-based Cross-domain Framework for Effective Hand Gesture Recognition

Cross-domain gesture recognition poses a significant challenge due to variations in sensing modalities, such as RGB and thermal imagery. To address this issue, we propose a domain adversarial neural network architecture that integrates a Convolutional Neural Network-based feature extractor with a Triplet Attention module. Using adversarial learning, our approach aligns feature distributions across domains, enabling robust classification of gestures regardless of modality. Experimental results indicate that the Triplet Attention module substantially enhances domain-invariant feature extraction, yielding high classification accuracy in both the source and target domains. By systematically evaluating multiple attention mechanisms, we confirm that Triplet Attention offers superior performance, improving recognition accuracy and resilience to domain shifts. Consequently, this method provides a strong foundation for practical applications that require adaptability to varying imaging conditions. It also opens avenues for broader cross-modal learning strategies that aim to enhance gesture recognition in diverse environments.