This paper presents a co-salient object detection method to find commonsalient regions in a set of images. We utilize deep saliency networks totransfer co-saliency prior knowledge and better capture high-level semanticinformation, and the resulting initial co-saliency maps are enhanced by seedpropagation steps over an integrated graph. The deep saliency networks aretrained in a supervised manner to avoid online weakly supervised learning andexploit them not only to extract high-level features but also to produce bothintra- and inter-image saliency maps. Through a refinement step, the initialco-saliency maps can uniformly highlight co-salient regions and locate accurateobject boundaries. To handle input image groups inconsistent in size, wepropose to pool multi-regional descriptors including both within-segment andwithin-group information. In addition, the integrated multilayer graph isconstructed to find the regions that the previous steps may not detect by seedpropagation with low-level descriptors. In this work, we utilize the usefulcomplementary components of high-, low-level information, and severallearning-based steps. Our experiments have demonstrated that the proposedapproach outperforms comparable co-saliency detection methods on widely usedpublic databases and can also be directly applied to co-segmentation tasks.
展开▼
