Evaluation and modeling of depth feature incorporated visual attention for salient object segmentation

Abstract

Visual attention has been used widely, such as region of interest (ROI) based image compression, imitating fixation region and salient object segmentation. The output saliency map range from spotlight map to object-based map which is related to different applications. We compare previously proposed saliency models and propose a new depth feature incorporated saliency model focusing on suppression of background saliency through piecewise function. We also produce an object-contour based ground truth database in order to evaluate several depth feature incorporated saliency models. Our method outperforms existing depth feature combination methods on the precision rate, when evaluated using the ground truth. At the same time, the effectiveness of using the depth feature in assisting salient object segmentation is verified.