Related Work - Hierarchical Neural Networks for Image Interpretation - page 64

upper part of the figure two vertical bar segments are grouped together, this grouping
is prevented in the lower part of the figure by a horizontal line of distractors, which
are grouped horizontally. In each case, grouping creates illusory contours. These
contours form the basis for surface-oriented computations, such as the filling-in of
color.
The vertical interactions also mediate attentional effects. Figure 3.21(b) shows,
how top-down spatial attention and bottom-up visual stimuli are integrated in layer
2/3 of V1. In this layer, the isolated collinear line-segments are grouped together.
Attention flows along this illusory contour and biases the entire object.
Many perceptual effects have been modeled with variants of this architecture.
The model is biologically plausible and it suggests micro-modules that could be
repeated to model higher visual areas. On the other hand, the model's architecture
is rather complex and it remains open, how the system performs when confronted
with natural visual stimuli.
3.3 Conclusions
A review of related work can never be comprehensive. Many approaches to image
interpretation exist in the literature that were not covered because the focus of this
chapter was to make the reader familiar to the concepts of hierarchy and recurrence,
which are central to the thesis.
While many models describe isolated aspects of human visual performance on
different levels of abstraction, so far no model is available that is biologically plau-
sible, involves horizontal and vertical recurrent interactions, and can be adapted
efficiently to perform different visual tasks.
Thus, there is clearly a need for further research. Neurobiology needs to find
out details of the neural circuitry that leads to the impressive performance of the
human visual system. Computational neuroscience must produce generic models
that capture the essential mechanisms without unnecessary detail. Psychophysics
can investigate properties of the visual system predicted by these models to test
them. Computer vision finally has the possibility to transfer these models to real-
world applications to validate their utility.