On Hierarchical Models for Visual Recognition and Learning of Objects, Scenes, and Activities

In many computer vision applications, objects have to be learned and recognized in images or image sequences. Most of these objects have a hierarchical structure.For example, 3d objects can be decomposed into object parts, and object parts, in turn, into geometric primitives. Furthermore, scenes are composed of objects. And also activities or behaviors can be divided hierarchically into actions, these into individual movements, etc. Hierarchical models are therefore ideally suited for the representation of a wide range of objects used in applications such as object recognition, human pose estimation, or activity recognition. In this work new probabilistic hierarchical models are presented that allow an efficient representation of multiple objects of different categories, scales, rotations, and views. The idea is to exploit similarities between objects, object parts or actions and movements in order to share calculations and avoid redundant information. We will introduce online and offline learning methods, which enable to create efficient hierarchies based on small or large training datasets, in which poses or articulated structures are given by instances. Furthermore, we present inference approaches for fast and robust detection. These new approaches combine the idea of compositional and similarity hierarchies and overcome limitations of previous methods. They will be used in an unified hierarchical framework spatially for object recognition as well as spatiotemporally for activity recognition. The unified generic hierarchical framework allows us to apply the proposed models in different projects. Besides classical object recognition it is used for detection of human poses in a project for gait analysis. The activity detection is used in a project for the design of environments for ageing, to identify activities and behavior patterns in smart homes. In a project for parking spot detection using an intelligent vehicle, the proposed approaches are used to hierarchically model the environment of the vehicle for an efficient and robust interpretation of the scene in real-time.