A user-centric approach to dataset augmentation for anomaly detection using Unity

Jansen, Marco ; Rosbak-Mortensen, Alexander ; Tøt, Mikkel Bjørndahl Kristensen ; Muhlig, Morten

4. term

Medialogy, Master

2023

2023-05-24

11 pages

Anomaly detection plays a critical role in surveillance systems, particularly in the automation of large-scale monitoring scenarios. Anomaly detection algorithms require datasets comprised of large amounts of annotated data to train and evaluate models. Gathering and annotating this data is a labour intensive task, that can be costly if outsourced to external partners. A method to get large amounts of annotated data explored by current state-of-the-art research is to generate it artificially using 3D applications, which has the advantage of being able to generate new frames in quick succession. This poses new issues for the end-user by being a specialized field, which means low-expertise users rely on external partnerships to have this option. In this paper we propose an application that synthesizes datasets using 3D models and simulates anomalies on real backgrounds using the Unity Engine. Additionally, we introduce a high-usability User Interface attached to a highly customizable simulation that simplifies the process of generating synthetic data without the need for specialized expertise in 3D animation. Testing datasets augmented with synthetic data made using our application gave promising results, with increases in both AUC and F1 scores in all cases. This indicates that synthetic data generation for low-expertise end users is a viable approach, and we recommend future works to focus on creating high variation in their data and to use photorealistic 3D models and lighting.

Download
View record in AAU Student Projects