Semi-supervised and Active Image Clustering with Pairwise Constraints from Humans

Clustering images has been an interesting problem for computer vision and machine learning researchers for many years. However as the number of categories increases, image clustering becomes extremely hard and is not possible to use for many practical applications. Researchers have proposed several methods that use semi-supervision from humans to improve clustering. Constrained clustering, where users indicate whether an image pair belong to the same category or not, is a well-known paradigm for semi-supervision. Past research has shown that pairwise constraints have the potential to significantly improve clustering performance. There are two major components to constrained clustering research: how pairwise constraints can be used to improve clustering (e.g: constrained clustering algorithms, distance or metric learning methods) and determining which constraints are most useful for improving clustering (e.g.: active or interactive clustering methods). In this thesis we propose three different approaches to improve pairwise constrained clustering spanning both of these components. First, we propose a distance learning method in non-vector spaces, where the triangle inequality is used to propagate the pairwise constraints to the unsupervised image pairs. This approach can work with any pairwise distance and does not require any vector representation of images. Second, we propose an algorithm for active image pair selection. A novel method is developed to choose the most useful pairs to show a person, obtaining constraints that improve clustering. Third, we study how pairwise constraints can effectively be used to cluster large image datasets. Complete clustering of large datasets requires an extremely large number of pairwise constraints and may not be feasible in practice. We propose a new algorithm to cluster a subset of the images only (we call this subclustering), which will produce a few examples from each class. Subclustering will produce smaller but purer clusters and can be used for summarization, category discovery, browsing, image search, etc.... Finally, we make use of human input in an active subclustering algorithm to further improve results. We perform experiments on several real world datasets such as faces, leaves, videos and scenes and empirically show that our approaches can advance the state-of-the-art in clustering