Knowledge-based view control of a neural 3-D object recognition system

The recognition of 3D objects is one of the most challenging goals in computer vision. In this paper we present a robot vision system that analyzes its environment by active vision techniques. Therefore, the system is gathering information about an object in the scene by taking multiple views. It is designed as a hybrid system that brings together the advantages of neural networks and knowledge based strategies. While the analysis of a single view is done by artificial neural networks, a knowledge based control determines viewpoints from which the scene should be analyzed in detail. Semantic networks are used to model 3D objects by holistically recognizable computer structures and by its basic views, thus, merging a part/part-of-hierarchy and an aspect-hierarchy. The system is implemented in conjunction with a six degrees of freedom robot and a hand-mounted camera.

[1]  G. Hartmann,et al.  How to provide pattern representations for learning and recognizing differently located objects from arbitrary camera positions , 1993, Proceedings of 1993 International Conference on Neural Networks (IJCNN-93-Nagoya, Japan).

[2]  Georg Hartmann,et al.  A multiple-view approach to 3D recognition based on complex model neurons1 , 1995 .

[3]  Georg Hartmann,et al.  Building an Artificial Retina for Distance- and Orientation-Invariant Pattern Recognition , 1993 .

[4]  G. Hartmann,et al.  Synchronization-Based Complex Model Neurons· , 1994 .

[5]  Gerhard Sagerer,et al.  Controlling Multiple Neural Nets with Semantic Networks , 1994 .

[6]  Robert Bergevin,et al.  Generic Object Recognition: Building and Matching Coarse Descriptions from Line Drawings , 1993, IEEE Trans. Pattern Anal. Mach. Intell..

[7]  Allen M. Waxman,et al.  Adaptive 3-D Object Recognition from Multiple Views , 1992, IEEE Trans. Pattern Anal. Mach. Intell..

[8]  Enrico Grosso,et al.  Relative positioning with uncalibrated cameras , 1992 .

[9]  O. Faugeras.  Stratification of three-dimensional vision: projective, affine, and metric representations , 1995 .

[10]  Charles R. Dyer,et al.  Model-based recognition in robot vision , 1986, CSUR.

[11]  H H Bülthoff,et al.  How are three-dimensional objects represented in the brain? , 1994, Cerebral cortex.

[12]  D. Hubel,et al.  Receptive fields of single neurones in the cat's striate cortex , 1959, The Journal of physiology.

[13]  Stephen Grossberg,et al.  Fast-learning VIEWNET architectures for recognizing three-dimensional objects from multiple two-dimensional views , 1995, Neural Networks.

[14]  Shimon Edelman,et al.  Representation of similarity as a goal of early visual processing , 1995 .

[15]  S. Edelman,et al.  Orientation dependence in the recognition of familiar and novel views of three-dimensional objects , 1992, Vision Research.

[16]  Georg Hartmann,et al.  Integrating Semantic and Neural Networks for Active Robot Vision , 1995 .

[17]  Linda G. Shapiro,et al.  Analysis of Scenes Containing Multiple Non-Polyhedral 3D Objects , 1995, ICIAP.