Fig 1.
The framework of our proposed study.
It consist of two parts, analysis part and proposed age-adapted saliency model as shown in figure, 29 years is mean age of adult observers. The sample image 1 is same for both of the models and it is only representation of the original image.
Table 1.
Saliency benchmark dataset.
Fig 2.
The process of generating human fixation map and human saliency maps of an image for age groups. OBn stands for the nth observer of an age group. The images used in this figure are similar but not identical to the original image, and is therefore for illustrative purposes only.
Fig 3.
Agreement analysis visualization.
The heat map visualizes agreement behavior in predicting the target fixation points by source saliency map and the ROC calculates the quantitative value of the agreement score. (a) Intra-age group: target fixation points of 4 years by source saliency map of 4 years. (b) Intra-age group: target fixation points of adults by source saliency map of adults. (c) Inter-age group: target fixation points of 4 years by source saliency map of adults. (d) Inter-age group: target fixation points of adults by source saliency map of 4 years.
Fig 4.
(a) The histogram of entropy indicates that there is a shift from left to right for 4 year to adult age group. (b) Entropy plotted in sorted order for different age groups over all the stimuli.
Fig 5.
Agreement analysis between observers within age group and across age groups.
(a) Intra-age group agreement scores, which reflects that kids agree more in explored locations than younger adults. (b) Agreement analysis results for the source saliency map of 4 years and adults in finding the target fixation of different age groups.
Table 2.
Average agreement score of human saliency map of observers from the source group in predicting fixation points of target age group.
Fig 6.
Age-based changes in center bias tendency across age groups.
Table 3.
Average prediction accuracy (AUC-score) by multi-scale feature subset selection.
Where as a∼b means from scale a to scale b are selected (Proposed S). Scale 6 is the coarsest level and scale 1 is the finest.
Table 4.
Average prediction accuracy by combining scale based subset selection, nonlinear integration and age-adapted center bias (Proposed S+I+C).
Fig 7.
Comparison of age-adapted proposed saliency models with baseline models of computational attention system.
Table 5.
Average prediction accuracy by proposed patch based method, scale 1, 2, 3, 4 are correspond to a, 16, 8 patch sizes respectively (Proposed P).
Table 6.
Comparison table of our proposed (S+I+C and P models) age-adapted models with available computational models of saliency prediction.