Email updates

Keep up to date with the latest news and content from BMC Neuroscience and BioMed Central.

This article is part of the supplement: Eighteenth Annual Computational Neuroscience Meeting: CNS*2009

Open Access Poster presentation

Grouping variables in an underdetermined system for invariant object recognition

Junmei Zhu* and Christoph von der Malsburg

Author Affiliations

Frankfurt Institute for Advanced Studies, 60438 Frankfurt am Main, Germany

For all author emails, please log on.

BMC Neuroscience 2009, 10(Suppl 1):P308  doi:10.1186/1471-2202-10-S1-P308


The electronic version of this article is the complete one and can be found online at: http://www.biomedcentral.com/1471-2202/10/S1/P308


Published:13 July 2009

© 2009 Zhu and Malsburg; licensee BioMed Central Ltd.

Introduction

We study the problem of object recognition invariant to transformations, such as translation, rotation and scale. A system is underdetermined if its degrees of freedom (number of possible transformations and potential objects) exceed the available information (image size). The regularization theory solves this problem by adding constraints [1]. It is unclear what constraints biological systems use. We suggest that rather than seeking constraints, an underdetermined system can make decisions based on available information by grouping its variables. We propose a dynamical system as a minimum system for invariant recognition to demonstrate this strategy.

A dynamical system for invariant recognition

Assume there are q objects in the gallery, and p possible transformations. An input image I is generated by one of the objects through a transformation. The task is to recover the object and the transformation that generate I. The system variables are C = (c1,..., cp)T for transformation and D = (d1,..., dq)T for object selection. When p + q > n, where n is the size of the image, the system is underdetermined, having many solutions.

Our system structure is shown in Figure 1. The state variables C and D follow the dynamics described by a system of linear differential equations. Figure 2 top row shows a solution of a toy system (n = 8*8, p = 72, q = 2), with I generated by c1 = 1, d1 = 1. In this underdetermined system, the solution depends on the initial condition.

thumbnailFigure 1. System architecture.

thumbnailFigure 2. Solutions (C and D) in a toy system.

The system can be made overdetermined by grouping variables such that all variables within a group share the same dynamics. When the total activity of the system is below a predefined level, we then let the variables in the top group resume their individual dynamics. Under this dynamics with grouping, the solution to the same toy system is shown in Figure 2 bottom row. It is close to the true value.

Discussion

Our example shows that, in an underdetermined system for invariant recognition, it is plausible to recover a sparse solution by grouping variables and then fine-tune the winning group. The applicability of this strategy depends on the structure of transformations and of objects. Our system could provide a model system to study the coarse-to-fine processing which is evident in biological systems [2].

Acknowledgements

Supported by EU project "SECO" and the Hertie Foundation.

References

  1. Poggio T, Koch C: Ill-posed problems in early vision: From computational theory to analog networks.

    Proceedings of the Royal Society London B 1985, 226:303-323. Publisher Full Text OpenURL

  2. Hegdé J: Time course of visual perception: Coarse-to-fine processing and beyond.

    Progress in Neurobiology 2008, 84:405-439. PubMed Abstract | Publisher Full Text OpenURL