Towards Fixation Prediction: A Nonparametric Estimation-Based Approach through Key-Points

Abstract

When we look at our environment, we primarily pay attention to visually distinctive objects. Saliency maps are topographical maps of the visually salient parts of scenes in which such visually distinctive objects, henceforth called visually important or salient, can be easily highlighted. Computing these maps is still an open problem whose interest is growing in computer vision. Thus, in this work, we propose a new method to compute these maps based on salient points extracted through local descriptors. After, a nonparametric kernel density estimation method is employed to estimate the final saliency map. In order to assess the performance, we carry out experiments on two large benchmark databases to demonstrate the proposed method performance against the state-of-the-art methods using different scoring metrics. Due to the experimental results obtained, we consider the proposed method is a valid alternative for saliency detection.