The lateral occipital complex is a large area of the brain that is diffusely active
during object recognition. Using fMRI, Kourtzi and Kanwisher show object selective
activation in the LOC, and demonstrate through fatigue effects that cells in the LOC
respond to structural (edge-based) properties [17]. Although their study can’t
determine what the structural properties are, Kosslyn and others [15] have suggested
they could be non-accidental properties of the type proposed by Lowe [19] and
Biederman [1]. Examples include edge collinearity, parallelism, symmetry and antisymmetry.
Psychological studies show that line drawings with non-accidental
features obscured are harder to recognize than obscured line drawings with nonaccidental
features intact [1].
This work models the LOC as computing fixed length non-accidental feature
transforms. The first and simplest example is the Hough transform – it projects edge
responses into the space of geometric lines, thereby making collinearity explicit. As
long as the temptation to threshold the Hough space and produce symbolic lines is
avoided, the Hough space is an appropriate feature representation for appearancebased
recognition. We are currently developing new transforms to capture other nonaccidental
features, such as parallelism, symmetry and anti-symmetry. The
preliminary results in this paper, however, show the surprisingly powerful results of
modeling the LOC as a Hough transform.
|
|
