Visual Search and Detection in Natural Scenes

点击翻译按钮获取中文摘要

Project Summary/Abstract The broad aim is to understand the mechanisms that underlie the ability of human and non-human primates to locate and identify target objects under natural conditions, a ubiquitous and fundamental visual task. Localization and identification of objects under natural conditions is difficult because there are so many dimensions of uncertainty (variability) in natural tasks: highly-variable properties of the background scenes in which the target object may be embedded, uncertainty about possible 2D (or 3D) location of the target object in the scene, uncertainty about the 2D (or 3D) orientation of the target object, and uncertainty about the contrast of the target object. These natural dimensions of uncertainty drove the design of the visual system (though evolution and learning). Thus, the theoretical and experimental analysis of the effects of uncertainty is critical for understanding the neural computations performed by the visual system, as well as the underlying neural circuits. In preliminary studies, we made substantial progress in developing the theory of identification and localization under natural conditions, in developing relevant computational and neurophysiological tools, in making relevant behavioral measurements in humans, and in making relevant physiological measurements in behaving monkeys. The first aim will be directed at rigorously testing a number of hypotheses about how various dimensions of target and background uncertainty singly, and in combination, affect behavior and neural computation. The second aim will be directed at measuring and modeling an important factor (partial masking) that affects detection and identification performance in natural backgrounds, but has never been systematically investigated in either simple or natural backgrounds. The third aim builds on the first two aims and will be directed a developing and testing a new theory of overt and covert visual search that takes into account natural scene statistics, variation in ganglion cell sampling and RF size with retinal location, and variation in intrinsic position uncertainty with retinal location. The theory encompasses many specific models that each makes detailed quantitative and qualitative predictions for arbitrary specific search stimuli. Covert and overt search tasks in noise and natural backgrounds will be used to discriminate between models.