View article

Conventionally, visual saliency means the most unique or distinct region or object of an image [1, 2, 3]; co-saliency detection aims at finding the common salient objects on multiple images [4, 5, 6]. However, beyond such distinctness and commonality, we argue that a determining factor underling saliency and co-saliency is the distributions of objects at the semantic level, which we define as the occurrence frequency 1 of each object in the image (s)(see Fig. 1 (i) and (j)). For both problems, the object distribution could be divided into two modes: the higher mode corresponds to high-frequent objects while the lower one to low-frequent objects. In the single-image case, high-frequent objects (eg, colorful flowers in Fig. 1 (a)) are usually taken as the background while the low-frequent ones (eg, basket, white and yellow cats in Fig. 1 (a)) are salient objects, which is consistent with the uniqueness assumption of …