Slide - Journal of Vision

From: Exact feature probabilities in images with occlusion
Journal of Vision. 2010;10(14):42. doi:10.1167/10.14.42
Figure Legend:
Joint feature probabilities log P(f 1, f 2) for wavelet pairs f 1 and f 2. For orthogonal, colocalized Haar wavelets (inset of
(A), shifted for visibility), the contours of the empirically sampled bivariate distribution are diamond-shaped for both (A)
natural images and (B) dead leaves images. At high feature amplitudes, certain object configurations have the greatest
likelihood and thus dominate the joint distribution. Panel (C) illustrates one such configuration. Colors indicate different
objects with unspecified intensities. Dark and light shadings show how the two wavelets weight the image pixels. (D)
Specifying only the object geometry (but not the object intensities) gives conditional feature distributions P(f 1, f 2∣ ) that
are of
bivariate
Gaussians
contours.
For the
conditional
distributions
that dominate
at high©feature
amplitude,
Date
download:
7/31/2017with ellipticalThe
Association
for Research
in Vision
and Ophthalmology
Copyright
2017. All
rights reserved.
the contours trace a diamond-shaped envelope (thick curve) as a function of the relative angles between the object