Reviews: Attention in Convolutional LSTM for Gesture Recognition

SUMMARY: This paper proposes a pipeline combining Res3D-3DCNN, convLSTM, MobileNet-CNN hybrid architecture for performing gesture recognition. In particular, it explores the integration of pooling and neural attention mechanisms in ConvLSTM cells. Four convLSTM variants are compared, which place pooling and attention mechanisms at different locations inside the cell. The pooling leads to a somewhat novel LSTM/convLSTM hybrid architecture. Finally, an attention-inspired gating mechanism is proposed, with some differences to the formulation in [5].