A Study on Eye Fixation Prediction and Salient Object Detection in Supervised Saliency. (2017)

• Record Type: Journal Article
• Authors: Johnson, Solwin; Subha, T.D.
• Is Part Of: Materials today. Volume 4:Number 2(2017)Part B
• Page Start: 4169
• Page End: 4181
• Publication Date: 2017
• Subjects: Feature selection -- Eye fixation prediction -- Salient object detection -- Saliency map; Materials science -- Congresses -- Periodicals; 620.1
• DOI: 10.1016/j.matpr.2017.02.119
• Languages: English
• ISSNs: 2214-7853

Abstract

Abstract: Based on human fixation data on natural images, there is an increasing demand in studying mappings from features to saliency maps. These models have achieved eclipse results than practically unsupervised saliency models. However, they constantly use more set of features annoying to explain all usable saliency-related factors, which leaves the truly effective features unknown and increases time cost. Through the supervised feature selection, it discloses that the features handled in existing models are highly redundant. On each of three touchstone datasets proposed in this paper is to predict the fixations of human eye with certain fixations method, some few numbers of features are found to be valuable enough. By using this model, the output is achieved in well manner with the help of the state-of-the-art. Both the features engaged and the model trained on any of the dataset exhibit good performance on the other two datasets indicating robustness of the occupied features and models across different datasets. At last, after training on a dataset for two different tasks, the prediction of eye fixation and detection of salient object is occupied and selected features prove robustness across the two tasks. So these findings suggest that a few set of features would account for visual saliency.