| This work overcomes some current limitations in approaches to automatic scene interpretation based on coarse-to-fine classification and extends this framework to object tracking. The main focus is to build computationally efficient systems that are invariant to changes in object appearance, addressing one primary cause for a lack of robustness in many computer vision applications, namely sensitivity to imaging variations in the environment. We propose a method to achieve such invariance at the level of individual classifiers by normalization using non-discriminating features. More globally, computational efficiency is achieved using a hierarchy of classifiers with coarse-to-fine search: most image regions are quickly and accurately rejected as background and processing is concentrated only on regions containing objects and object-like structures. Instead of manually determining the structure of the hierarchy as in previous work, we propose a method to automatically learn the architecture and the set of classifiers from data. We also introduce a stochastic model to account for the interactions amongst the different classifiers in a hierarchy, which improves discrimination and unifies object detection and tracking within a single generative framework. |
