Luka Fürst (2007) Izbira značilnic za detekcijo objektov dane vizualne kategorije na slikah. MSc thesis.
This master's thesis deals with the problem of selecting features for automatic detection of objects of a given visual category in images. The goal of object detection is to determine the locations and sizes of all objects of the given category in each input image, where the number of objects displayed in individual images is not known in advance. Various circumstances may contribute to the difficulty of object detection, such as cluttered backgrounds, diversity in the appearance of objects in the category, diversity in the scale of the displayed objects, and partial occlusions. The thesis primarily focuses on systems that represent each input image by a set of features and use the resulting image representations to form detection hypotheses. The feature set is assumed to be prepared by a two-step procedure in the training stage. In the first step, the initial feature set is extracted from a set of training images. In the second step, the features to be used in the test stage are selected from the initial set. The thesis is devoted to the second step, i.e., feature selection. An effective selection method reduces the computational complexity in the test stage and eliminates features that are useless or even harmful for detecting objects of a given category. In the master’s thesis, we present two feature selection methods. At the core of the first one (the so-called filter method) is a feature evaluation function that is based on a transformation of the problem of detecting objects in training images into a problem of classifying training images themselves. Since much existing theory and practice of feature selection pertains to the classification rather than detection domain, such a problem transformation greatly expands the range of applicable selection approaches. Because of its design, the feature evaluation function can be straightforwardly incorporated into the AdaBoost selection framework, which selects features by considering their interdependence in an implicit manner. The second selection method that is presented in this thesis (the so-called wrapper method) systematically runs the detection procedure on a fixed training image set for various candidate feature sets. The goal of such a selection approach is to find a feature set that enables the detection procedure to attain the highest detection accuracy on the training image set. The problem of finding the optimal candidate feature set can be easily transformed into a problem of finding the optimal vertex in the corresponding state space graph. In the implementation of the system, each of the two selection methods was integrated with a well-known detection approach and with two promising state-of-the-art feature extraction methods. We experimentally tested all four combinations of feature extraction and feature selection methods. The experiments were evaluated using five test image sets, each of which defined a visual category and an independent detection setup. In some cases, as it turned out, a very small number of selected features suffices to bring about a fairly high detection accuracy. Particularly remarkable results were achieved for the combination of the filter selection method, the feature extraction method due to Fidler et al. (2006), and the UIUC car dataset. In this setup, the selection of merely four features out of 952 initial candidates led to a detection accuracy that is comparable to the state-of-the-art results for the UIUC dataset.
Actions (login required)