Efficient Approaches for Object Class Detection
M. Villamizar
Advisors: Alberto Sanfeliu, Juan Andrade Cettto and Francesc Moreno Noguer

Abstract - Computer vision and more specifically object recognition have demonstrated in recent years an impressive progress that has led to the emergence of new and useful technologies that facilitate daily activities and improve some industrial processes. Currently, we can find algorithms for object recognition in computers, video cameras, mobile phones, tablets or websites, for the accomplishment of specific tasks such as face detection, gesture and scene recognition, detection of pedestrians, augmented reality, etc. However, these applications are still open problems that each year receive more attention in the computer vision community. This is demonstrated by the fact that hundreds of articles addressing these problems are published in international conferences and journals annually. In a broader view, recent work attempts to improve the performance of classifiers, to face new and more challenging problems of detection and to increase the computational efficiency of the resulting algorithms in order to be implemented commercially in diverse electronic devices. Although nowadays there are robust and reliable approaches for detecting objects, most of these methods have a high computational cost that make impossible their application for real-time tasks. In particular, the computational cost and performance of any recognition system is determined by the type of features, the method of recognition and the methodology used for localizing objects within images. The main objective of these methods is to produce not only effective but also efficient detection systems. Through this dissertation different approaches are presented for addressing efficiently and discriminatively the detection of objects in diverse and difficult imaging conditions. Each one of the proposed approaches are especially designed and focus on different detection problems, such as object categorization, detection under rotations in the plane or the detection of objects from multiple views. The proposed methods combine several ideas and techniques for obtaining object detectors that are both highly discriminative and efficient. This is demonstrated experimentally in several state-of-the-art databases where our results are competitive with other recent and successful methods. In particular, this dissertation studies and develops fast features, learning algorithms, methods for reducing the computational cost of the classifiers and integral image representations for speeding up feature computation.




Presentation Slides