How far we away from a perfect visual saliency detection-DUT-OMRON: a new benchmark dataset

Visual saliency detection has gained more and more attentions from academic and industrial researchers in the last 3 or 4 years. Due to great efforts being put into this field, many recent proposed algorithms have very good evaluation results on existing datasets. However, we argue that visual saliency detection is still far away from perfect because such good results are mainly due to the simplicity and bias of existing datasets. In this paper, we propose a new DUT-OMRON dataset which, to our best knowledge, is the first visual saliency detection dataset that has both the bounding box and eye fixations ground-truth in large scale. We evaluated 14 state-of-the-art methods on the proposed dataset, and the accuracy curves on proposed dataset are much lower than that on existing datasets. We believe our dataset is more challenging than existing ones and therefore leave more space for researchers to improve their algorithms.