Better together: Fusing visual saliency methods for retrieving perceptually-similar images

Abstract

In this paper, we describe a new model of visual saliency by fusing results from existing saliency methods. We first briefly survey existing saliency models, and justify the fusion methods as they take advantage of the strengths of all existing works. Initial experiments indicate that the fused saliency methods generate results closer to the ground-truth than the original methods alone. We apply our method to content-based image retrieval, leveraging a fusion method as a feature extractor. We perform experimental evaluation and show a marked improvement in retrieval performance using our fusion method over individual saliency models.