Fig. 5: Fixation frequency for object categories.
From: Eye movements during free viewing to maximize scene understanding
a Bounding boxes for objects in each Winograd pair were manually created, and the object categories were assigned to them. The scene understanding relevant (SU-relevant) and the SU-irrelevant categories have the same set of objects across each Winograd pair. b Fixation frequencies based on the object categories in free viewing and scene description conditions across 25 participants per image (and 36 images) show that the SU-relevant object was most fixated (p = 0.005 for comparison of fixation frequency to SU-relevant and DeepGaze categories in free viewing; p < 0.001 for all other comparisons). The search target was most fixated in the object search condition (p < 0.001 for all comparisons). The SU-relevant object was fixated on the most for the counting objects, although the difference relative to the SU-irrelevant was smaller (p = 0.029). c ΔCFR,I: The difference in cumulative fixation frequency observed (for 25 participants per image, 36 images) between SU-relevant and SU-irrelevant object categories was highest in the scene description condition (different from free viewing at 6th fixation onward, p = 0.043), followed by free viewing (different from counting objects at 4th fixation onward, p = 0.049). No significant difference was observed across any fixation for the other two conditions. (****p < 0.0001; ***p < 0.001; **p < 0.01; *p < 0.05). A one-tailed bootstrapped analysis was conducted to test the significance of all the results. All analyses had their significance levels corrected for the False Discovery Rate (28 comparisons, 20 shown here). In b, the box plot shows a participant-level distribution of fixation frequencies, and the line within the box indicates the median. The box spans the interquartile range (IQR), and the whiskers extend to the most extreme values within 1.5 × IQR. In c, the central measure is the mean, and the error bars show the 68% bootstrap confidence interval. Consent was obtained from the person in the featured image for its publication.
