Large-Scale Self-Supervised Human Activity Recognition

In this paper, a self-supervised approach is used to obtain an effective human activity representation using a limited set of annotated data. This research is aimed on acquiring human activity representation in order to improve the accuracy of classifying videos of human activities in the NTU RGB+D 120 dataset. The effectiveness of various self-supervised approaches, as well as a supervised method, is studied. The results reveal that when the training set gets smaller, the performance of supervised learning approaches diminishes, whereas self-supervised methods maintain their performance by utilizing unlabeled data.