Exploring model-based target discrimination metrics
Gary Witus, Marshall R. Weathersby · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 2004
Visual target discrimination has occurred when the observer can say "I see a target THERE!" and can designate the target location. Target discrimination occurs when a perceived shape is sufficiently similar one or more of the instances the observer has been trained on. Marr defined vision as "knowing what is where by seeing." Knowing "what" requires prior knowledge. Target discrimination requires model-based visual processing. Model-based signature metrics attempt to answer the question "to what extent does the target in the image resemble a training image?" Model-based signature metrics attempt to represent the effects of high-level top-down visual cognition, in addition to low-level bottom-up effects. Recent advances in realistic 3D target rendering and computer-vision object recognition have made model-based signature metrics more practical. The human visual system almost certainly does NOT use the same processing algorithms as computer vision object recognition, but some processing elements and the overall effects are similar. It remains to be determined whether model-based metrics explain the variance in human performance. The purpose of this paper is to explain and illustrate the model-based approach to signature metrics.