While bag-of-words (BoW) scene descriptor has been widely used for scene retrieval applications, the BoW descriptor alone often fails to capture local details of a scene and produces poor results. In this paper, we address this issue by a simple effective approach, “un-supervised salient part discovery”, in which a set of salient parts are discovered via scene parsing and used as additional queries for the scene retrieval. Further, we also address the issue of discovering salient parts in a scene, and present a solution that provides similar parts for similar scenes. Multiple ranking results from the individual part queries are then integrated into a final ranking result by adopting an unsupervised rank fusion technique. Experimental results using challenging scene dataset validate the effectiveness of our approach.
展开▼
