Salient object detection via point-to-set metric learning

Abstract

Distance metric is an essential step of salient object detection, in which the pairwise distances are often used to distinguish salient image elements (pixels and regions) from background elements. Instead of using the point-to-point distance metrics which possibly implicitly take into account the context information around data points, we learn the point-to-set metric to explicitly compute the distances of single points to sets of correlated points and cast saliency estimation as the problem of point-to-set classification. First, we generate a series of bounding box proposals and region proposals for an input image (i.e., some pre-detected regions which possibly include object instances), and exploit them to compute a recall-preference saliency map and a precision-preference one, based on which the background and foreground seed regions are respectively determined. Next, we collect positive and negative samples (include point samples and set samples) to learn the point-to-set distance metric, and employ it to classify the image elements into foreground and background classes. Last, we update the training samples and refine the classification result. The proposed approach is evaluated on three large publicly available datasets with pixel accurate annotations. Extensive experiments clearly demonstrate the superiority of the proposed approach over the state-of-the-art approaches.