Two-Stage Co-Salient Object Detection

Most of the existing methods achieve co-saliency detection at single level. In this work, we combine the object-level and region-level processing to detect co-salient objects in a group of images. At the object level, we formulate proposal selection as an outlier detection problem. We find good region proposals and generate a template for each image. At the region level, we introduce the smoothness constrain to present a multi-constraint classification model and use the templates from an image group to self-train a group-specific classifier, thereby predicting the saliency labels for each superpixel. Experimental results show that the proposed method outperforms other state-of-the-art co-saliency methods.