Global-Local Enhancement Network for NMF-Aware Sign Language Recognition

Sign language recognition (SLR) is a challenging problem, involving complex manual features (i.e., hand gestures) and fine-grained non-manual (NMFs) facial expression, mouth shapes, etc .). Although are dominant, also play an important role in the expression of sign word. Specifically, many words convey different meanings due to features, even though they share same gestures. This ambiguity introduces great challenges words. To tackle above issue, we propose simple yet effective architecture called Global-Local Enhancement Network (GLE-Net), including two mutually promoted streams toward crucial aspects SLR. Of streams, one captures global contextual relationship, while other stream discriminative cues. Moreover, lack datasets explicitly focusing on this kind feature, introduce first non-manual-feature-aware isolated Chinese dataset (NMFs-CSL) with total vocabulary size 1,067 daily life. Extensive experiments NMFs-CSL SLR500 demonstrate effectiveness our method.