Object Class Detection Using Part-Based Models Trained from Synthetically Generated Images
Supervisor(s) and Committee member(s): Rainer Lienhart (advisor), Bernhard Möller (committee member), Eckehard Steinbach (committee member)
This thesis presents part-based approaches to object class detection in single 2D images, relying on pre-built CAD models as a source of synthetic training data. Part-based models, representing an object class as a deformable constellation of object parts, have demonstrated state-of-the-art results with respect to object class detection. Typically, the majority of part-based approaches rely on real training images of publicly available image data sets and consequently, the positive output of those detectors is restricted to the viewpoints which are represented by those real training images. However, progress in the domain of computer graphics enables the generation of photo-realistic renderings on demand from a database of CAD models, which can serve as training source for learning an object class detection approach. In this thesis, we present part-based object class detection methods which are based on synthetically generated positive training images and real negative training images, thereby combining the advantages of the two domains described above. More specifically, photo-realistic object parts, representing the object class being trained, are learnt in an unsupervised way without requiring any manual bounding box, object part, or viewpoint annotations during the training process. The established object parts are efficiently combined into an object class detection framework relying on two part-based models with different learning paradigms. In addition, we outline an extension of our detection framework which is able to cope with multiple object classes. The approaches to object class detection are evaluated on standard benchmark data sets and achieve state-of-the-art results with respect to object class detection in single 2D images.