Cite as:
Sharan L, Adelson E H, Motoyoshi I, Nishida S, 2007, "Non-oriented filters are better than oriented filters for skewness detection" Perception 36 ECVP Abstract Supplement

Non-oriented filters are better than oriented filters for skewness detection

L Sharan, E H Adelson, I Motoyoshi, S Nishida

A material that has mesostructure, such as stucco, looks darker and glossier when the skewness of the luminance histogram is higher. Skewness is a useful cue, and humans seem to use it. We have proposed a simple computational model for estimating skewness with neurons, in which the ON and OFF output streams of filters are each run through a nonlinearity, followed by pooling and differencing (Motoyoshi et al, 2007 Nature 447 206 - 209). We now show computationally that non-oriented (center - surround) filters have advantages over oriented (Gabor-like) filters in this task. We used the skewness of filter outputs (ie sub-band histograms) to classify images as having positive or negative luminance skewness, and estimated d'. Non-oriented filters typically outperform oriented filters by factors ranging from 2 to 10; this is true both for natural images and for random noise. We also find no psychophysical evidence for orientation selectivity in the skewness aftereffect (Motoyoshi et al, 2007 Nature loco cit). Thus, non-oriented filters seem to dominate skewness computations.

[Supported by NTT CS Labs, NSF.]

These web-based abstracts are provided for ease of seaching and access, but certain aspects (such as as mathematics) may not appear in their optimum form. For the final published version of this abstract, please see
ECVP 2007 Abstract Supplement (complete) size: 2091 Kb