An Experimental Analysis of Deep Learning Models for Human Activity Recognition with Synthetic Data

In this paper, an experimental study of state-of-the-art techniques in Human Activity Recognition (HAR) is presented. Different Deep Learning algorithms, including CNNs and RNNs, are examined and compared. The experimental part is done using two real-life datasets Kinetics-400 and UCF-101 and one synthetic - SURREACT. All of them are used both for training and testing. The models - SlowFast, X3D and MViT are evaluated using accuracy top-1, and the results are identifying the best-performing combinations of dataset and model. One important question this study is trying to answer is whether a synthetic dataset can replace a real-world one. Finally, limitations and future directions are discussed, along with potential real-world applications.