Few-Shot Training of Prototype Networks for Sign Language Recognition

Limited proficiency in sign language creates communication barriers, motivating the development of robust Automatic Sign Language Recognition (SLR) systems. We address isolated SLR in a low-resource setting using few-shot metric-based meta-learning. Sign videos are encoded with spatiotemporal convolutional backbones and classified using a prototypical network, enabling generalization to unseen classes from small support sets. We compare the SlowFast architecture with state-of-the-art video models on the LSA64 benchmark under strict class-disjoint protocols. SlowFast achieves 94.33% accuracy, outperforming competing backbones and demonstrating an effective and data-efficient approach for low-resource isolated SLR.