PUB - Publikationen an der Universität Bielefeld

Computational models of visual saliency have been used to detect salient regions and simulate human eye-gaze on images and videos. A majority of the existing approaches are highly parametric in nature. They are specialized to predict either eye-gaze or detect salient regions, but not both simultaneously. Like other computer vision approaches, the saliency models too impose pre-specified grids to process the image. In this context we explore ways of exploiting random/ stochastic algorithmic approaches for saliency computation to address issues like pre-specified grids, computational efficiency, parameter set etc. We propose three different approaches for saliency computation on images and provide elaborate benchmarking results with respect to other saliency systems. Consequently, we have been successful in improving the state-of-the-art in terms of eye-gaze prediction and salient region detection performance of the saliency systems. In addition, we have extended one of our proposed saliency approaches to predict eye-gaze while viewing a tutoring or goaldirected action scenario. Along with the proposed algorithms, we also have created a video dataset for evaluating saliency systems in the context of goal-directed action. We hope that the proposed approaches for saliency computation, exprimental protocols, resulting video dataset and the ensuing discussions will help the community in developing more sophisticated systems of visual saliency.