Features and Matching - Computer Vision for Visual Effects - page 138

was found to perform the best for lighting and scale changes. In general, far fewer
matches were found when comparing images of 3D objects versus images of planar
scenes.

4.3.1 Learning Good Parameters

Based on the success of SIFT and GLOH, Winder, Brown, and Hua [ 548 , 549 , 71 ]
undertook an extensive effort to learn good parameters for the general class of
histogram-based descriptors. They began with a very large set of known match and
non-match pairs of patches obtained from DoG interest points, and simultaneously
learned the parameters and combinations of descriptor components that gave the
best performance on the training set. For example, they considered the number of
gradient orientationbins, grid configurationof SIFTandGLOHregions, configuration
and Gaussian variance of DAISY regions, and number of basis vectors for dimension-
ality reduction using PCA, among many other factors. They generally recommended
that a DAISY-like configuration was best, where the Gaussian variances increase with
distance from the feature location. These studies differed from the testing procedure
of Mikolajczyk et al. in that the training and testing data were based on ground-truth
point correspondences in general images of large 3D scenes acquired by multi-view
stereo techniques (see Section 8.3 ).

4.4 COLOR DETECTORS AND DESCRIPTORS

It's somewhat surprising that almost all research on detectors and descriptors
assumes a grayscale image as input. Since distinctive color regions provide obvious
cues for featuredetection that areoften lost or subduedby conversion tograyscale, it's
worthwhile to investigate detectors and descriptors that preserve color information
throughout the process. Here we mention a few such algorithms.
In terms of color detectors, Kenney et al. [ 237 ] extended the Harris matrix from
Section 4.1.1 using several basic axioms for properties of a good corner detector.
They showed how to generalize the Harris matrix in Equation ( 4.3 ) and Shi-Tomasi
criterion in Equation ( 4.13 ) to images with multidimensional range (such as color
images) and/or domain (such as 3Dmedical images). Unnikrishnan andHebert [ 505 ]
introduced a generalization of scale space for feature detection in color images. They
proposed a family of functions involving first and second derivatives of the three color
channels at a given scale, so that the output of the function at each scale is roughly
invariant to linear illumination changes. Similar to Harris-Laplace, feature points
are detected at local extrema in scale space. Forssén [ 148 ] described an extension of
MSERs to color. The one-dimensional watershed algorithm to compute the extremal
regions is replaced with an agglomerative clustering scheme to produce the nested
set of connected components.
In terms of color descriptors, early work focused on invariant-based techniques.
For example, grayscale moment invariants were generalized to color images by
Mindru et al. [ 330 ]. That is, we compute the generalized moments
x m y n

r g b

( I R ( x , y )) ( I G ( x , y )) ( I B ( x , y )) (4.44)