= PDF Reprint, = BibTeX entry, = Online Abstract
A. Borji, D. N. Sihite, L. Itti, Quantifying the relative influence of photographer bias and viewing strategy on scene viewing, In: Proc. Vision Science Society Annual Meeting (VSS11), May 2011. (Cited by 4)
Abstract: Saccade distributions while observers freely watch natural scenes and videos are often found to be highly biased toward the image center (center-bias effect) (Tatler, 2007). Our quantitative comparison of 30 saliency models over three standard datasets of still images (Bruce & Tsotsos 2006, Kootstra et al., 2008 and Judd et al., 2009), shows that model rankings do not agree. Interestingly, a trivial central Gaussian blob saliency model outperforms many models in regard to predicting where humans look. Two main sources of center-bias are: photographer bias (natural tendency of photographers to place objects of interest near the center) and viewing strategy (tendency of subjects to look at the center to extract more information) (Tseng et al., 2009). In this study, we measure the relative influence of these causes and introduce a less center-biased dataset as a benchmark for fair evaluation of models. From four datasets (three aforementioned and Le Meur et al., 2006), we chose those images with the lowest center-bias index (a defined measure of tendency of human saccade density maps to be concentrated toward center) and selected just 187 out of overall 1250 stimuli. The average center-bias index of accepted images, all original stimuli and Gaussian blob were 0.61, 0.76 (0.88 for Judd) and 1, respectively. Next, to remove the variability in eye recording parameters in datasets, we recorded eye movements of 30 subjects watching these images. The center-bias index for recorded eye movements over selected images and the Judd dataset were 0.76 and 0.861, respectively. After removing the first saccade, these values dropped to 0.68 and to 0.845. Although selected images had less objects at the center, there was still a great amount of saccade density at the center. Our results suggest that, 1) Widely used datasets are center-biased, 2) Photographer bias could be reduced, and 3) Viewing strategy has a higher influence than photographer bias on overall center-bias since removing photographer bias does not reduce overall center-bias significantly.
Themes: Human Psychophysics, Computational Modeling, Human Eye-Tracking Research
Copyright © 2000-2007 by the University of Southern California, iLab and Prof. Laurent Itti.
This page generated by bibTOhtml on Thu Jan 31 11:39:41 PST 2019