A Comparative Study of Saliency Aggregation for Salient Object Detection

Abstract

A variety of saliency detection methods have been proposed in recently, which often complement each other. In this study, we try to improve their performances by aggregating these individual ones. First, we propose an improved Bayes aggregation method with double thresholds. Then, we compare it with five other aggregation approaches on four benchmark datasets. Experiments show that all the aggregation methods significantly outperform each individual one. Among these aggregation methods, average and Non-negative Matrix Factorization (NMF) weights perform best in terms of precision-recall curve, our Bayes is very close to them. While for mean absolute error score, NMF and our Bayes perform best. We also find that it is possible to further improve their performance by using more accurate reference map. The ideal is ground truth, of course. Our results could have an important impact for applications required robust and uniform saliency maps.