Special Issue on Bio-inspired Vision Systems

Abstract

Listen

Translate article

Much of early vision research drew inspirations from what was then known about biological vision. Progress was slow however, and over time machine learning in combination with features selected more on pragmatic grounds took over. Increasingly impressive results seem to justify that approach. The advent of affordable depth cameras further moved the field (certainly in robotics) away from biological considerations—why bother how the human brain arrives at 3D scene interpretations when the 3D data is just readily available? Not all problems simply vanish however by throwing novel sensors and heavy machine learning at them. 3D sensors really only give 2.5D data. The backside, as well as sensor artefacts coming from physical limitations, still need to be filled in. And how meaning can be attached to visual percepts—2D or 3D—can not simply be explained by learning from large hand labelled data bases. So there is still a lot to learn from biological vision systems. How to arrive at a sufficiently clear (whatever that means in detail) interpretation of the scene from several patchy cues? How to tightly couple vision to other aspects of a cognitive system? What is the right level of abstraction for representations? In this issue we present current work in bio-inspired vision systems and explore the possibilities offered by new findings in biological vision systems as well as latest developments in machine vision. The survey article by Kruger et al. starts with a short history of biologically motivated methods in computer vision and lists several open problems in current computer vision research. This is contrasted with key findings from the primate’s visual system, and the article goes on to argue for rethinking the potential impact of biologically motivated methods in the light of these new findings. One of these findings is the importance of shared intermediate level representations. Rodriguez-Sanchez et al. discuss in their article the role such intermediate-level representations play in achieving higher-level object abstraction and present recent developments in the neural computational modeling of intermediate-level shape processing. Another important finding is that biological vision seems to maintain several possibly conflicting (think: Necker Cube) interpretations of a visual scene and avoid early commitment to a single solution. This is explored in M. Zillich (&) Automation and Control Institute, Vienna University of Technology, Vienna, Austria e-mail: zillich@acin.tuwien.ac.at