Compute lambdas for channel power law scaling.
For a broad family of features, including gradient histograms and all
channel types tested, the feature responses computed at a single scale
can be used to approximate feature responses at nearby scales. The
approximation is accurate at least within an entire scale octave. For
details and to understand why this unexpected result holds, please see:
P. Dollár, R. Appel, S. Belongie and P. Perona
"Fast Feature Pyramids for Object Detection", PAMI 2014.
This function computes channels at multiple image scales and plots the
resulting power law scaling. The purpose of this function is two-fold:
(1) compute lambdas for fast approximate channel computation for use in
chnsPyramid() and (2) provide a visualization of the power law channel
scaling described in the BMVC2010 paper.
chnsScaling() takes two main inputs: the parameters for computing image
channels (pChns), and an image or set of images (Is). The images are
cropped to the dimension of the smallest image for simplicity of
computing the lambdas (and fairly high resolution images are best). The
computed lambdas will depend on the channel parameters (e.g. how much
smoothing is performed), but given enough images (>1000) the computed
lambdas should not depend on the exact images used.
USAGE
[lambdas,as,scales,fs] = chnsScaling( pChns, Is, [show] )
INPUTS
pChns - parameters for creating channels (see chnsCompute.m)
Is - [nImages x 1] cell array of images (nImages may be 1)
show - [1] figure in which to display results
OUTPUTS
lambdas - [nTypes x 1] computed lambdas
as - [nTypes x 1] computed y-intercepts
scales - [nScales x 1] vector of actual scales used
fs - [nImages x nScales x nTypes] array of feature means
EXAMPLE
sDir = 'data/Inria/train/neg/';
Is = fevalImages( @(x) {x}, {}, sDir, 'I', 'png', 0, 200 );
p = chnsCompute(); lambdas = chnsScaling( p, Is, 1 );
See also chnsCompute, chnsPyramid, fevalImages
Piotr's Computer Vision Matlab Toolbox Version 3.25
Copyright 2014 Piotr Dollar & Ron Appel. [pdollar-at-gmail.com]
Licensed under the Simplified BSD License [see external/bsd.txt]