GEOGENX: Generic parts-based object detection from satellite imagery

Expanding constellations of earth observation satellites are drastically increasing the availability of satellite imagery to the US government, providing timely, detailed, and readily available intelligence for mission planning and situational awareness on a global scale. Under the DARPA funded SBIR Phase 2 GEOMETRIX program, VSI developed an innovative ecosystem to distill and reconcile commercial satellite imagery with other noisy, unreliable, and often conflicting multi-source inputs supported by state-of-the-art machine learning and computer vision techniques. \n\n The GEOMETRIX program provided clear indicators for continued Phase II research and development. A critical component of the system is object detection, useful to both confirm the existence of reported targets as well as extend analysis to uncover previously unreported events. Unfortunately, detection performance degrades against more challenging imagery with varying viewpoint, sun angle, resolution, and atmospheric effects. Further, object labels are taken from a limited taxonomy (e.g., “aircraft” rather than “MiG‑29”) that requires additional analyst intervention to label true targets of interest. \n\n VSI thus proposes the sequential SBIR Phase II GEOGENX program, an innovative computer vision capability to locate, recognize, and mensurate a deep taxonomy of object types through both generic parts‑based object detection as well as geometric, geographic, and semantic constraints. VSI is well suited to program challenges, offering expertise in remote sensing, object detection, 3D reconstruction, and machine learning; extensive experience in the development of innovative software solutions for challenging defense and intelligence problems; and a thorough understanding of the needs of image analysts across the defense and intelligence community.