Local features.
The key question is which local features to use, namely how to decrease λb while
maintaining relatively high ‘on object’ probabilities po. The first possibility would
be to use the oriented edges which serve as input to the entire system. These are
the most basic local features detected in the visual system. The problem with
edges however is that they appear with rather high density in typical scenes. In
other words λb is such that λb|B| > 1, leading to very large numbers of false
positives. See also detailed discussion in Grimson (1990). The edge detector used
here is a simple local maximum of the gradient with 4 orientations, horizontal
and vertical of both polarities, see figure 1. They exhibit strong invariance to
photometric transformations and local geometric transformations. Typical λb for
such edges on generic images is .03 per pixel.
We therefore move on to flexible edge conjunctions to reduce λb while maintaining
high po. These conjunctions are on one hand flexible enough to be stable at
specific locations on the object, for the previously defined range of deformations.
On the other hand they are lower density in the background. These conjunctions