Reducing the Feature Distance between Real and Synthetic Data for Training Safeguards-Relevant Computer Vision Models

Computer vision models have great potential as tools for international nuclear safeguards verification activities. Computer vision applications could help safeguards inspectors and analysts more quickly interpret visual scenes, identify objects of interest, or notice abnormal occurrences. However, open-source examples of some safeguards-relevant objects are rare, making training computer vision models difficult. Synthetic data is a potential method to supplement the lack of available training data, but prior work training models on synthetic data and testing them on real data has demonstrated a performance gap, in which models have low performance. To address this low performance, we explore methods to characterize – and subsequently reduce – the feature differences between real and synthetic data. We then evaluate the performance of models trained and tested on data sets that are closer together for selected feature spaces. In this paper, we discuss the methods we have applied to bring the data distributions closer together and describe impacts on model performance. Finally, we will contrast our findings across different computer vision approaches.