Measuring detection and localization performance

Abstract

This paper proposes a simple, but general, model for describing an observer's ability to find and report target objects in grayscale image backgrounds, which provides a basis for measuring the observer's combined detection- and-localization performance in various image-interpretation tasks. The model assumes that: 1) the observer's image detection response and first-choice of target location both depend on the “maximally suspicious” finding on the image, 2) a correct (first-choice) localization of the actual target occurs if and only if its location was regarded as the most suspicious and 3) a target's presence does not alter the suspicion engendered by any other (normal) image findings. Formalization of these assumptions relates the ROC curve, which measures the observer's ability to discriminate between target and nontarget images, to the “Localization Response” (LROC) curve, which measures the conjoint ability to detect and correctly localize the actual targets in the images. A two-parameter “binormal” version of this model provides the basis for an iterative, maximum-likelihood procedure that concurrently fits both the ROC and LROC curves, using an observer's image-ratings and target-localizations for a set of image interpretations. That same model can be extended to multiple-report (“free-response”) interpretations and to multiple-target images, provided that the observer's detection capability and criterion for reporting possible targets both remain invariant across images and across successive reports on a given image. That extension of the model leads to formulations for the so-called “Free-Response” (FROC) curve, and for a recently proposed “Alternative FROC” (AFROC) curve.