A structure-based disentangled network with contrastive regularization for sign language recognition

Liqing Gao, Lei Zhu, Lianyu Hu, Peng Shi, Liang Wang, Wei Feng

Abstract: Highlights•We propose to structurally recognize signs to improve the generality of the model.•We design a structure-based disentangled model with three jointly optimized branches.•A SDM to decouple complete features into entity and template features.•A TFM to aggregate complementary knowledge from three branches.•A TCM to establish contextual relationships between sign language actions.