Multi-Hypothesis Automated Wireframe Generation

High-accuracy standoff missile targeting systems such as JASSM require prominent geometric features for automatic target correlation. Three-dimensional wireframe models can precisely describe object boundaries (edges and corners) and are thus well-suitedrepresentations of fixed targets. The past two decades of computer vision research have seen considerable advances in wireframe construction from stereoscopic imagery, but significant limitations still exist in today's state-of-the-art systems: interactivemodel construction is laborious and time intensive, while automated systems create numerous spurious objects and incorrect hypotheses due to physical phenomena such as reflectance, shadows, and occlusion. Our Phase I effort demonstrated a model-based approach that discriminates among competing wireframe hypotheses. These hypotheses are constructed using image features such as lines and corners, as well as by applying combinations of physicalRobust metrics compare the original stereo imagery with synthetic prediction images generated by each model, incorporating visibility information such as shadows and obscuration. This hypothesis evaluation guides the refinement of parameters such as roofheight to improve accuracy. Our Phase II effort will extend the Phase I technologies and create an end-to-end prototype system for automated wireframe generation for subsequent transition into existing application frameworks.