Hierarchical Graphical Models

Abstract

The goal of the proposed hierarchical graphical models is to recognize various instances of different object classes in images, image sequences or other scene representations like e.g. occupancy grid maps. The term “object” in this context is used as a general term representing visual objects, visual parts, visual features, visual primitives, but also activities, actions or motion primitives. In the following we will regard two different kinds of hierarchies:A compositional hierarchy and a similarity hierarchy. In compositional hierarchies the structure of a parent node is defined by its children, where edges define the spatial or spatiotemporal relation between the parent and the children nodes. In this manner complex high-level nodes can be recursively defined based on simple low-level features. Similarity hierarchies, on the other hand, describe similarities among objects, and among parts. In this work, they will be combined with a coarse-to-fine search by means of scale space representation. They are used to increase the robustness of the representation as well as the overall runtime performance.