Exploring Domain Randomization's Effect on Synthetic Data for Activity Detection

The construction and operation of Metaverse virtual environments, e.g., 3D reconstruction and activity detection is an important supporting technology of computer vision. Recently, synthetic data has seen a surge in adoption for model training in computer vision. Prior research generally show a positive correlation between the volume of synthetic training data and inference accuracy. This paper focuses on the domain of activity detection, and explores how to improve the performance of such algorithms using synthetic data. In particular, we present an overview of the state-of-the-art in using domain randomization approaches for synthetic data generation. This paper presents initial inference accuracies of a model trained on initial attempts at domain randomized synthetic data (7.2%), compared to a model trained on real-world data (9.2%). The synthetic data, although performed worse, indicated promising trajectories for future work, approximately 2% away from the real-world result.