Abstract
When an observer looks at an image, his eyes fixate on a few select points that correspond to interesting image locations. However, how this process is affected by image resolution is not well understood. Here we investigate how image resolution affects human fixations through an eye tracking experiment. We showed 100 images at different resolutions to 30 observers. Each image was shown at one of seven resolutions (width of 4, 8, 16, 32, 64, 256, 1024 pixels) and upsampled to the original size of 1024x768 pixels for display. We found that: 1) As image resolution decreases, users fixate in fewer locations and fixations get more concentrated near the center. 2) Fixations from lower resolution images can predict fixations on higher resolution images. We measure how well one observer's fixations predict another observer's fixations on the same image at different resolutions using the area under the ROC curves as a metric. Fixations on an image at full resolution are predicted better by fixations on the same image as the resolution increases, but the rate of improvement declines after a resolution of 64px. 3) Fixations are most consistent across users on images at a resolution of 32 and 64px. More specifically, the fixations on 32 and 64 resolution images predict fixations on the same image better than fixations from any other resolution predict fixations of that resolution. The fixations on the lowest and highest resolution images are harder to predict. These findings suggest that working with fixations at an image resolution of 32-64px could be both perceptually adequate and computationally attractive.
NSF CAREER awards 0447561 and IIS 0747120. Frédo Durand acknowledges a Microsoft Research New Faculty Fellowship, a Sloan Fellowship, Royal Dutch Shell, the Quanta T-Party, and the MIT-Singapore GAMBIT lab.