Abstract
While there has been much research investigating visual recognition memory, little work has examined the specific content of visual free recall memory, despite evidence that these may be two neurally distinct processes. Here, we quantify the capacity and resolution of visual free recall for complex real-world scene images using a drawing task. Participants (N=30) studied 30 real-world scenes (10s each) and after an 11-min digit span distractor task, drew as many images as possible from memory in as much detail as possible. To serve as benchmarks, separate participants made 1) drawings from the scene category names ("lower bound"), reflecting canonical representations of a scene, and 2) drawings created while looking at the image ("upper bound"), reflecting the maximum information that could be drawn. We leveraged online crowd-sourced experiments on Amazon Mechanical Turk to objectively score the content of these 1,782 drawings. These ratings revealed an impressive detail contained in peoples' memories. First, memory drawings were easily matched to their corresponding image, and were nearly as diagnostic as those drawn directly from the image. Second, memory drawings contained 73.9% of the objects in the drawings made directly from the image, and on average, participants recalled 151.3 objects across the experiment with very few false alarms (1.83 objects across the experiment). Third, the spatial arrangement of objects in memory drawings was highly accurate, and almost identical to the original image. Further, computer vision graph-based visual saliency maps significantly predicted which objects would be remembered by participants. Collectively these results suggest that visual recall memory contains diagnostic, detailed, and precise representations of real-world scenes.
Meeting abstract presented at VSS 2018