FAST-DATA: Fully Automated Synthetic Training Data for Accelerated Training of AI

The primary goal of this effort is to create a system for generating high-quality synthetic datasets that can be used to effectively, precisely, and rapidly train AI/ML models designed to detect threats on a battlefield. This system will allow us to significantly reduce the time between the emergence of new threats and the deployment of effective AI/ML-based countermeasure, as well as produce smaller and more effective AI/ML models.áIn this Phase 1 effort we will focus on AI/ML models which process visual inputs from cameras to identify, track, and characterize threats. Specifically, we will be focused on UAVs (drones) as the threat. We will create and combine 3D models of these UAVs with landscape video and environmental conditions to produce the synthetic training data. UAVs were selected as the initial use case as they are evolving very quickly in current conflicts. They also serve as a good proxy for other vehicles, personnel, and physical threats that can be detected visually. Upon successful completion of this Phase 1 effort, we can expand to other signals, such as audio and EM data.áThis project will produce synthetic datasets in 4 increasingly challenging ways, starting with the simplest caseáin which 3D UAV models are produced from 2D images, and the background environmental conditions are unmodified/neutral. The second iteration will involve datasets generated with varying environmental conditions (such as lighting conditions or image quality). The final two iterations (3 and 4) will attempt to address the issue of producing synthetic training data when no video or images of the UAVs are available (e.g., prior to them being observed and recorded on the battlefield). Specifically, iteration 3 will produce the 3D UAV models from schematics, and iteration 4 will attempt to produce the 3D UAV models from verbal descriptions, aided by generative AI tools.áThe performance will be measured in each iteration by a comparison between our in-house AI/ML UAV detection system when trained on the synthetic datasets, versus the same system trained on real-world UAV video and still images.áOur team brings substantial relevant experience, including deep AI/ML expertise and the development of a low-cost, low-power UAV detection and tracking system that will serve as the performance evaluation baseline. This existing system, operating on visual and audio inputs from commodity hardware, demonstrates the team's capability in the problem domain.áThis approach could greatly enhance our ability to produce high quality synthetic AI/ML training datasets. This would not only allow AI/ML models to be trained more quickly, but also more precisely as the training data could be tuned to specific deployment environments or conditions. This should produce smaller models which would require fewer resources to run during the inference phase, and would also be more effective than larger, more general, models.