Synthetic data-seeded active learning for automated ATR data labeling

Data labeling is often the most time consuming and expensive component of new automatic target recognition (ATR) model development in remote sensing. Synthetic data have long been desired for ATR model development in scenarios with limited real data available but have not been fully adopted in defense due to compute constraints and truth limitations within critical applications. This work combines AgileView’s synthetic data generation platform with Bedrock Research’s remote sensing foundation models to perform automated active learning. We seed initial model training on an example use case of maritime object classification solely with synthetic data, and then an iterative feedback loop of foundation model fine tuning with exclusively real data fully automates the active learning process and minimizes the human hours required for comprehensive data labeling. Furthermore, this technique introduces novel measures of label and image stability to automatically quantify the reliability of an auto-generated label. This provides a fast and trustworthy approach to ATR data and model curation for widespread defense applications.