ZEUS: Zero-Shot Explainable Universal Segmentation-Based ATR (ZEUS-ATR)

Historical computer vision approaches applied to automatic target recognition (ATR) require massive amounts of data because they are trying to solve a 3-D problem directly in 2-D representations. Novel foundational models developed on internet-scale databases enable the direct estimation of scene segmentations and entity range from single and multiple monocular images. These 3-D data representations provide a powerful new space for performing ATR. By acting directly on the 3-D representations of the world, next-gen ATR algorithms can directly integrate, e.g., target CAD models for training instead of massive data collections with novel sensors. Furthermore, operating in the 3-D representation enables principled multi-sensor fusion, and enables the exact same target detection and classification algorithms to process data from disparate sensing modalities (e.g., RGB, IR, and SAR). This proposal describes CoVar’s Zero-Shot Explainable Universal Segmentation-Based (ZEUS) ATR a novel approach to cross-modal foundational model training, and 3-D world-space informed ATR. Once built, ZEUS will have zero new data training requirements and enable the same exact ATR models to operate on disparate sensing data.