Publication

Abstract : An analytical conclusion based on eye tracking data sets has shown that Graph Based Visual Saliency (GBVS) measures saliency in a better way. GBVS promotes higher saliency at the center of the image plane and strongly highlights salient regions even for the locations that are far-away from object borders. It predicts human fixations more consistently than the standard algorithms. Every pixel in an image is mapped as an individual graph node in the activation map. This in turn increases the computational time. Hence the objective of this paper is to analyze the performance of saliency measure in GBVS by modeling different grouping strategies to represent a node. Here, we concentrate on finding the dissimilarity between the nodes by grouping pixels as a node with overlapping or non-overlapping pixels in the surrounding nodes which optimize the saliency closer to the Eye-Tracker’s saliency. The different grouping strategies of GBVS are analyzed across several performance measures like Normalized Scanpath Saliency the Linear Correlation Coefficient, Area Under Curve, , Similarity, Kullback – Leibler Divergence to prove its efficiency. Key terms – Visual Attention Models, Saliency maps, Eye-Tracking, Grouping pixels.