Appearence Based Object Representation for Stereo Vision

Research project P 13981 Appearance Based Object Representation For Stereo Vision Robert SABLATNIG 24.01.2000 3D object recognition and pose estimation constitutes the basic step in more complex tasks like tracking or manipulator control. In conventional setups, 3D object recognition requires the availability of a user-defined, geometrical (CAD) object model and/or reliable correspondences between model and image features. In the appearance based object representation approach, neither a user defined object model nor feature correspondences are required. Instead, the object model is built using raw intensity images of the object under various viewing and lighting conditions. These images can be efficiently compressed using subspace techniques like PCA; new instances of the object are then expressed as a linear combination of a small orthogonal set of basis images (the so called eigen-images). The coefficients of valid linear combinations define a multidimensional manifold within the subspace spanned by the eigen-images; this fact is used in so called parametric subspace methods to recover various object parameters, e.g., pose. Parametric subspace models have been used successfully in the past to solve various object recognition tasks, including tracking, effector positioning and visual inspection. Most of these systems, however, model only a small subset of the object`s appearance parameters, for example 2 dimensional translation and lighting. In the course of our project, we will develop a parametric subspace model that can represent arbitrary (i.e, 6 degrees of freedom) object poses. In particular, we will address the following problems: efficient representation of high dimensional parametric manifolds, development of subspace methods for stereo vision, robust retrieval, autonomous and incremental model (subspace) construction. In the course of the project, a software library for appearance basedobject representation shall be designed and implemented. Experiments including industrial tasks like visual inspection and positioning of an articulated robotic arm will be conducted in order to demonstrate the usefulness and practical relevance of our system.