A simple saliency model predicts a number of motion popout phenomena

Abstract

Visual search for a moving target among stationary distractors is more efficient than searching for a stationary target among moving distractors, and searching for a fast target among slow distractors is more efficient than vice versa. This indicates that the ease of search for a target with a particular motion is not determined simply by the difference between target and distractor velocities. We suggest a simple model for predicting ease of search for a unique motion, based upon a quantitative measure of target saliency. Essentially, search will be easier the more the target motion deviates from the general pattern of velocities in the scene. Our model predicts a number of well-known motion search phenomena, and suggests that one control for target saliency as well as target discriminability when drawing conclusions about visual system mechanisms from search experiments.

Visual saliency does not account for eye movements during visual search in real-world scenes

“…First, computational models have been developed that use known properties of the visual system to generate a saliency map or landscape of visual salience across an image (Itti & Koch, 2000Koch & Ullman, 1985). In these models, the visual properties present in an image give rise to a 2D map that explicitly marks regions that are different from their surround on image dimensions such as color, intensity, contrast, and edge orientation (Itti & Koch, 2000;Koch & Ullman, 1985;Parkhurst, Law, & Niebur, 2002;, contour junctions, termination of edges, stereo disparity, and shading (Koch & Ullman, 1985), and dynamic factors such as motion (Koch & Ullman, 1985;Rosenholtz, 1999). The maps are generated for each image dimension over multiple spatial scales and are then combined to create a single saliency map.…”

“…Pattern information is acquired only during periods of relative gaze stability (fixations) due to a combination of central suppression and visual masking (Matin, 1974;Thiele, Henning, Buishik, & Hoffman, 2002;Volkman, 1986). Gaze control is the process of directing the eyes through a scene in real time in the service of ongoing perceptual, cognitive, and behavioral activity (Henderson, 2003;Henderson & Hollingworth, 1998, 1999.…”

Learning top down scene context for visual attention modeling in natural images

“…All visual attention algorithms generate a saliency map with a predicted pixel-level saliency. This saliency map is thresholded at k=1, 3,5,10,15,20,25 and 30 percent to obtain binary saliency maps. The percentage of human fixations contained within each binary map is the performance measure.…”

“…Early visual attention models [9,10] are pure bottom up approaches and use multiple low level image features such as intensity, color, orientation, texture and motion to determine regions of interest in natural images. In these approaches feature specific saliency maps are computed for every low level feature and the final master map is a linear or non linear combination of individual feature specific saliency maps.…”

Free Choice in Deontic Inquisitive Semantics (DIS)

“…But let me first elaborate on the notion of salience that I would like to use, which is a notion of contextualized perceptual salience inspired by recent research on visual salience in terms of informativity or surprise (e.g. Rosenholtz, 1999;Itti and Koch, 2001; Bruce and Tsotsos, 2009; Itti and Baldi, 2009). The general idea is that, when presented with a scene, those things stand out that are unexpected.…”