Preview

PDF, English Download (10MB) | Terms of use

Abstract

The objective of this thesis is to enhance visual recognition for objects and scenes through the development of novel mid-level representations and appendent learning algorithms. In particular, this work is focusing on category level recognition which is still a very challenging and mainly unsolved task. One crucial component in visual recognition systems is the representation of objects and scenes. However, depending on the representation, suitable learning strategies need to be developed that make it possible to learn new categories automatically from training data. Therefore, the aim of this thesis is to extend low-level representations by mid-level representations and to develop suitable learning mechanisms. A popular kind of mid-level representations are higher order statistics such as self-similarity and co-occurrence statistics. While these descriptors are satisfying the demand for higher-level object representations, they are also exhibiting very large and ever increasing dimensionality. In this thesis a new object representation, based on curvature self-similarity, is suggested that goes beyond the currently popular approximation of objects using straight lines. However, like all descriptors using second order statistics, it also exhibits a high dimensionality. Although improving discriminability, the high dimensionality becomes a critical issue due to lack of generalization ability and curse of dimensionality. Given only a limited amount of training data, even sophisticated learning algorithms such as the popular kernel methods are not able to suppress noisy or superfluous dimensions of such high-dimensional data. Consequently, there is a natural need for feature selection when using present-day informative features and, particularly, curvature self-similarity. We therefore suggest an embedded feature selection method for support vector machines that reduces complexity and improves generalization capability of object models. The proposed curvature self-similarity representation is successfully integrated together with the embedded feature selection in a widely used state-of-the-art object detection framework. The influence of higher order statistics for category level object recognition, is further investigated by learning co-occurrences between foreground and background, to reduce the number of false detections. While the suggested curvature self-similarity descriptor is improving the model for more detailed description of the foreground, higher order statistics are now shown to be also suitable for explicitly modeling the background. This is of particular use for the popular chamfer matching technique, since it is prone to accidental matches in dense clutter. As clutter only interferes with the foreground model contour, we learn where to place the background contours with respect to the foreground object boundary. The co-occurrence of background contours is integrated into a max-margin framework. Thus the suggested approach combines the advantages of accurately detecting object parts via chamfer matching and the robustness of max-margin learning. While chamfer matching is very efficient technique for object detection, parts are only detected based on a simple distance measure. Contrary to that, mid-level parts and patches are explicitly trained to distinguish true positives in the foreground from false positives in the background. Due to the independence of mid-level patches and parts it is possible to train a large number of instance specific part classifiers. This is contrary to the current most powerful discriminative approaches that are typically only feasible for a small number of parts, as they are modeling the spatial dependencies between them. Due to their number, we cannot directly train a powerful classifier to combine all parts. Instead, parts are randomly grouped into fewer, overlapping compositions that are trained using a maximum-margin approach. In contrast to the common rationale of compositional approaches, we do not aim for semantically meaningful ensembles. Rather we seek randomized compositions that are discriminative and generalize over all instances of a category. Compositions are all combined by a non-linear decision function which is completing the powerful hierarchy of discriminative classifiers. In summary, this thesis is improving visual recognition of objects and scenes, by developing novel mid-level representations on top of different kinds of low-level representations. Furthermore, it investigates in the development of suitable learning algorithms, to deal with the new challenges that are arising form the novel object representations presented in this work.