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 lowlevel 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.