Abstract
Identification of classic attributes guiding search is based on experiments using unstructured displays. These attributes do not explain the efficiency of search in real scenes. We propose that “depth guidance” can massively reduce the effective set size in real scenes. Fast, non-selective processes are known to provide information about spatial layout and about proto-objects. Given approximate distances and object sizes in the image, only a very few proto-objects can possibly be the current target. To test, we had participants draw boxes on 200 real-world images (indoor and outdoor) indicating multiple possible locations and sizes for target objects (cats, cups, and people) that were not in the images. In the main experiment, one box was picked as the target. Other box locations were chosen as distractors and the boxes resized to match the target's image size. Thus, if observers were looking for cats, only one box would be the right size to just hide the cat. On each trial, observers used as few mouseclicks as possible to identify that box. At chance performance, the slope of the function relating guesses to set size is 0.5. On average, the guesses x set size slope was 0.30. If dots replaced boxes, eliminating the possibility of depth guidance, the slope increased to 0.52 (chance). Our depth guidance slope estimate of 0.30 is too steep because some distractors were chosen from boxes placed at the roughly the same distance as the target object, making those boxes of an appropriate size. These are scored as distractors but could be targets. An additional experiment estimated that this occurred on 42% of trials. Correcting the slope for this factor yields an estimated slope of 0.19. This shallow slope suggests that depth guidance effectively reduces the number of candidate targets in real scenes by directing attention to size-appropriate objects.
F32EY019815-01, ONR N000141010278.