SBIR/STTR Award attributes
CoVar’s DOCTRINAIRE is a new approach to computer aided object annotation that is modeled after the way expert end-users leverage generic, robust background information (e.g., what wheels look like) and known doctrine (the size and shape of components on a pickup truck) to perform reliable, explainable object detection and annotation. Our approach solves the robustness problem by training a reliable object-of-interest detection network. These detections are then fed to a separate segmentation network, which identifies each discrete part of an object on a pixel-by-pixel basis. The outputs of this part segmentation network are then integrated into a doctrine-based target identifier that can be updated using only target models. This enables DOCTRINAIRE to annotate new types of objects even if the system has never seen any sensor data for objects from these new classes. In this work, we propose to formalize a mathematical treatment of DOCTRINAIRE and extend current implementations to make it relevant to the automatic annotation of massive datasets of interest to NGA, commercial satellite, and aerial data sources.
