Parallel temporal feature selection based on improved attention mechanism for dynamic gesture recognition

Abstract Dynamic gesture recognition has become a new type of interaction to meet the needs daily interaction. It is most natural, easy operate, and intuitive, so it wide range applications. The accuracy depends on ability accurately learn short-term long-term spatiotemporal features gestures. Our work different from improving performance single network with convnets-based models recurrent neural network-based or serial stacking two heterogeneous networks, we proposed fusion architecture that can simultaneously gestures, which combined in parallel. At each stage feature learning, gestures are captured simultaneously, contribution networks classification results spatial channel axes be learned automatically by using attention mechanism. sequence pooling operation module compared through experiments. And proportion learning quantitatively analyzed, final model determined according experimental results. used for end-to-end method was validated EgoGesture, SKIG, IsoGD datasets got very competitive performance.