Semi‐supervised long short‐term memory for human action recognition

Abstract

In real human action recognition task, it is a common phenomenon that there are many unlabelled samples and few labelled samples. How to make good use of unlabelled samples to improve the generalisation ability of models is the focus of semi-supervised learning research. In this study, the authors present two semi-supervised methods based on long short-term memory (LSTM) to learn discriminative hidden features. One is the LSTM ladder network, the other is the Symmetrical LSTM network. By them unlabelled samples can be used automatically to improve learning performance without relying on external interaction. Both on the NTU-RGB + D dataset and the Kinetics dataset, their methods achieve >10 and 5% improvements, separately.