Evaluation on fusion of saliency and objectness for salient object segmentation
Linwei Ye, Zhi Liu, Lina Li · 2015
Saliency detection measures the probability how a region attracts human visual attention, and objectness estimates the probability that a rectangle window may contain potential objects. Can a salient object segmentation method which utilizes both saliency and objectness achieve a better segmentation performance? To address this problem, this paper evaluates different fusion schemes to integrate saliency with objectness for effective salient object segmentation. Based on the saliency map generated by any saliency model and the pixel-level objectness map, the resultant fusion map is exploited to initialize salient object and background. Then a fusion map based salient object segmentation method under the framework of graph cut is proposed to obtain the final salient object segmentation result. We performed extensive experiments on two public datasets, and conclude that fusion of saliency and objectness generally facilitates to improve salient object segmentation performance compared to only using saliency or objectness, and the proposed segmentation method using a number of fusion schemes with saliency models outperforms the state-of-the-art salient object segmentation method.