DOCTRINe-based AIded target REcognition (DOCTRINAIRE)

Modern AiTR approaches can be significantly improved by understanding how DoD training doctrine can be incorporated into the AiTR problem.  DoD training enables persons with limited a priori knowledge of different military actions to reliably classify potential threats based on some simple rules about the relationships between entity components (e.g., wheels, wings, rotors, windows). Most modern network architectures (YOLO, CenterNet, etc.) promise end-to-end detection and classification, but these networks exacerbate the problems of applying deep learning to military data (due to lack of training data, propensity to over-train, and inability to adapt to new classes). Although robust, reliable, end-to-end classification of military vehicles is difficult for single shot networks, deep learning can be used to robustly solve smaller sub-problems (entity detection, part segmentation) and these can be combined to solve the overall AiTR problem robustly.  By properly combining these capabilities into a coherent AiTR system, CoVar's DOCTRINAIRE approach enables robust, extensible, explainable AiTR that enables reliable target detection and target classification across widely varying backgrounds. Furthermore, DOCTRINAIRE systems are trivially adaptable to detecting new types of targets even when no data from those targets has been collected.