Saliency map model for visual attention in depth

Abstract

Previous saliency-based visual attention models dealt with the mechanism underlying attentional shifts in the early visual system. Such models can predict the distribution of gaze movements on the basis of saliency maps (Itti & Koch, 2000). Saliency maps are calculated on the basis of the distribution of color in and the orientation of input images. In recent studies, saliency-based models were modified to predict the gaze positions for moving objects. The data from psychophysical experiments indicate that singletons defined by depth easily pop out, indicating that the saliency of depth information is implicated in visual information processes. However, depth information is not considered in the calculations for saliency maps. In this study, we modified a saliency-based model with regard to using depth information for calculating the saliency map and deciding the gaze position. Our model can discriminate each object in the visual scene by referring to the region of shared depth and shift the gaze points from object to object on the basis of depth saliency. We analyzed the results for the visual search tasks in which some objects distributed to the depth space and obtained the scan order of objects by tracking the mouse cursor. The predicted scan orders for objects were found to be very similar to those obtained in the experimental results for the early selection phase. This indicates that the saliency of depth information may play an important role in visual search in depth space.