Abstract
Recent work by Isola et. al. (2011) has demonstrated that memorability is an intrinsic property of images that is consistent across viewers and can be predicted accurately with current computer vision techniques. Despite progress, a clear understanding of the specific components of an image that drive memorability is still unknown. While previous studies such as Khosla et. al. (2012) have tried to investigate this computationally, no behavioral study has systematically explored the memorability of image regions within individual images. Here we study which regions from an image are memorable or forgettable. Using a large image database, we obtained the memorability scores of the different visual regions present in every image. In our task, participants viewed a series of images, each of which were displayed for 1.4 seconds. After the sequence was complete, participants similarly viewed a series of image regions and were asked to indicate whether each region was seen in the earlier sequence of full images. A selection of these regions were taken from the sequence of images participants viewed previously while the others were taken from images never seen before. We then computed consistency ratings for the memorability of each region across participants. Results indicate the first direct evidence from subject data that memorability of specific image regions is consistent across observers (i.e. individual regions remembered by one viewer are remembered by others as well), and that certain regions within a particular image are more memorable than others. Furthermore, we show that this region memorability is different from bottom-up saliency by demonstrating a weak correlation between the two.
Meeting abstract presented at VSS 2015