PhD Thesis Semi-Supervised Ensemble Methods for Computer Vision

V isual object classification and tracking are two of the cardinal problems in computer vision. Both tasks are extremely complicated and far from being solved. Recent advances towards building better detection and tracking systems were mainly achieved by improved representations and applying better learning algorithms. For the learning, usually supervised algorithms are applied which demand large amounts of hand-labeled data in order to yield accurate results. However, hand-labeling is a tedious and timeconsuming task, which is also prone to human errors. Additionally, learning only from labeled data is not natural and there exist tasks, such as tracking-by-detection, where the learners have to be able to exploit both labeled and unlabeled data. Also, the growing number of digital images present in the web and off-line databases makes human handlabeling hardly possible. Hence, we need learning techniques that are able to exploit huge amounts of unlabeled data with a reduced quantity of human interaction. These considerations have led to an increased interest in semi-supervised learning methods that learn from a small amount of labeled data and a large amount of unlabeled data. In this thesis, we propose several novel approaches to semi-supervised learning using ensemble methods, such as boosting or random forests, and show their applicability on various computer vision tasks. The reasons for studying ensemble methods is that they are powerful and fast and are already used in many computer vision applications. In the first part of the thesis, we propose a novel semi-supervised boosting algorithm based on visual similarity learning and demonstrate its applicability to object detection. Then, we extend the method to on-line learning. In the second part of the thesis, we propose a novel random forest method that is able to learn from both labeled and unlabeled data. We demonstrate the benefits of the approach on both several machine learning tasks and object categorization where our method is able to train one inherent multi-class classifier rather than several one-versus-all classifiers. In the subsequent chapter, we demonstrate how to extend random forests to on-line mode. We also show how to apply random forests to a learning paradigm that is very similar to semi-supervised learning, i.e., multiple instance learning. Finally, we hypothesize that visual object tracking can be formulated as a oneshot-semi-supervised learning task. In the tracking experiments, we demonstrate that applying the proposed on-line semi-supervised and multiple instance learning methods lead to more stable and higher accurate tracking results.