User-Centered Pipeline for Synthetic Augmentation of Anomaly Detection Datasets

Abstract

Automatic anomaly detection plays a critical role in surveillance systems but requires datasets comprising large amounts of annotated data to train and evaluate models. Gathering and annotating these data is a labor-intensive task that can become costly. A way to circumvent this is to use synthetic data to augment anomalies directly into existing datasets. This far more diverse scenario can be created and come directly with annotations. This however also poses new issues for the computer-vision engineer and researcher end users, who are not readily familiar with 3D modeling, game development, or computer graphics methodologies and must rely on external specialists to use or tweak such pipelines. In this paper, we extend our previous work of an application that synthesizes dataset variations using 3D models and augments anomalies on real backgrounds using the Unity Engine. We developed a high-usability user interface for our application through a series of RITE experiments and evaluated the final product with the help of deep-learning specialists who provided positive feedback regarding its usability, accessibility, and user experience. Finally, we tested if the proposed solution can be used in the context of traffic surveillance by augmenting the train data from the challenging Street Scene dataset. We found that by using our synthetic data, we could achieve higher detection accuracy. We also propose the next steps to expand the proposed solution for better usability and render accuracy through the use of segmentation pre-processing.