Improving Saliency Models by Predicting Human Fixation Patches

Abstract
There is growing interest in studying the Human Visual System (HVS) to supplement and improve the performance of computer vision tasks. A major challenge for current visual saliency models is predicting saliency in cluttered scenes (i.e. high false positive rate). In this paper, we propose a fixation patch detector that predicts image patches that contain human fixations with high probability. Our proposed model detects sparse fixation patches with an accuracy of 84 % and eliminates non-fixation patches with an accuracy of 84 % demonstrating that low-level image features can indeed be used to short-list and identify human fixation patches. We then show how these detected fixation patches can be used as saliency priors for popular saliency models, thus, reducing false positives while maintaining true positives. Extensive experimental results show that our proposed approach allows state-of-the-art saliency methods to achieve better prediction performance on benchmark datasets.

Citation
Dubey, R., Dave, A., & Ghanem, B. (2015). Improving Saliency Models by Predicting Human Fixation Patches. Lecture Notes in Computer Science, 330–345. doi:10.1007/978-3-319-16811-1_22

Publisher
Springer Nature

Journal
Lecture Notes in Computer Science

Conference/Event Name
12th Asian Conference on Computer Vision, ACCV 2014

DOI
10.1007/978-3-319-16811-1_22

Additional Links
http://link.springer.com/chapter/10.1007%2F978-3-319-16811-1_22http://vcc.kaust.edu.sa/Documents/B. Ghanem/papers/fixation_patches_accv2014.pdf

Permanent link to this record