Geometric object models have been widely used for visual object tracking. In this contribution we present particle filter based object tracking with pose estimation using an appearance based lightfield object model. A light-field is an image-based object representation which can be used to render a photo realistic view of an arbitrarily shaped object from arbitrary viewpoints. It is shown how lightfield object models can be generated and utilized. Furthermore, we show how these models fit into the probabilistic framework of dynamic state estimation by defining an appropriate likelihood distribution from an image similarity metric. Finally, we present results and accuracy evaluations from tracking experiments of different objects. 1

this article, we propose a new kind of model for robot vision, the so called plenoptic model, a model that can be acquired fully automatically. This is regarded in this article as learning. The key idea is to collect a sequence of images from objects or scenes using a hand--held camera. These images together with some extra information, which is computed automatically during model building, serve as a model of the object or scene. Using this model photo--realistic views from objects and scenes can be created (rendered) synthetically, even if the view has not been observed before