Unsupervised foreground detection in unlabeled images

This paper presents an unsupervised method to simultaneously identify foreground objects and improve image clustering quality. Given a set of unlabeled images, each image is decomposed into a set of local features and feature weights. First, the unlabeled images are automatically clustered based on feature appearance similarity and geometry similarity. Then, feature weights are recomputed according to appearance and geometry correspondences within the clustering results. Finally, we iteratively refines the clustering results as well as feature weights which reflect the degree of features belong to foreground objects. These novel aspects lead to precise foreground feature identification and improvement in overall detection performance compared with previous methods in Caltech-101 dataset.