Award

Automatic target recognition (ATR) remains a challenge due to the large variety of Army relevant sensors, sensing modalities, imaging conditions, and environments, nuanced differences in military targets, and the wide range of image scales over which a target may be imaged. Labeling vast amounts of training data to cover all of these conditions and targets is not a feasible approach. With Hi-VALUE ATR Kitware and Johns Hopkins University will enable an analyst to efficiently create new object detectors with zero or only a few object exemplars in the Linchpin Unclassified Environment by describing the object types in plain, descriptive text in addition to providing a few visual examples. We use a hybrid architecture, combining state-of-the-art vision language models (VLMs) with classical algorithmic components (e.g. Gaussian mixture models for motion detection, inpainting for synthetic occlusion generation, feature matching and homography estimation for image stabilization). Hi-VALUE ATR’s detectors will create results without lengthy retraining and quickly adapt to new domains (environments, targets, sensors, sensor modalities, weather conditions, etc.) because we build on modern VLMs and state-of-the-art few-shot learning methods. Preliminary results indicate that our approach is highly effective at ATR while requiring only limited training data.