Interactive Deep Model Tuning for Surveillance
Review

Video Surveillance is a time-consuming task for nuclear safeguards inspectors, having to review the video footage from hundreds of cameras installed in nuclear facilities worldwide. In recent years, deep learning has proved to obtain outstanding results in common computer vision tasks that are relevant for safeguards inspectors such as object detection or scene understanding. Deep learning models have the potential to significantly improve the video review workflow both by providing automated data analysis and by enabling interactive tools supporting the manual surveillance review. However, several challenges need to be addressed in the context of nuclear safeguards. Use cases vary greatly between different facility types and therefore deep learning models need to be trained or fine-tuned to each specific task. Labelling large sets of training data for each model is not feasible, as it would require too much effort and cannot be outsourced due to the sensitivity of the data. We propose an interactive workflow using pre-trained models that integrates data review, labelling and model tuning. It minimizes the labelling and training effort and gives the inspector control over the tasks learned by the model. The paper describes the workflow and underlying model architecture and presents experimental results.