filt_get_sized.m

Standard model object recognition matlab code

function f = filt_get_sized (f, sizes, shift, tag)

% FUNCTION f = filt_get_sized (f, sizes, shift, tag)
% 
% Sometimes, the size (x-y dimension) of the filter may not be
% pre-determined or easy to calculate.  In such cases, this
% function produces a filter with a specified size.  The input
% arguement, sizes, has 3 numbers, corresponding x, y, and old
% feature dimensions.  For example, in the original model, the
% filter from S2 to C2 is ones with the same sizes as S2.  

% Maximum number of nonzero afferents.
max_nonzero_aff = sizes(1)*sizes(2);

if length(sizes) < 3
  num_new_feature = 1;
  sizes = [sizes 1];
else
  num_new_feature = sizes(3); % same as num_old_feature = sizes(3)
end

% One way of building filters: more intuitive, but when there are
% many features, one runs out of MEMORY. (because we are building
% 4-D filter matrix.
%
%f1 = zeros(sizes(1), sizes(2), sizes(3), sizes(3));
%filt_template = ones(sizes(1), sizes(2));
%for i = 1:num_new_feature
%  f1 (:,:,i,i) = filt_template;
%end
%f2 = f1;
%f = [];
%f = filt_package (f, f1, f2, shift, tag);

% This is an alternative method.  We package filters by hand, and
% avoid the MEMORY issue.
f_template = ones(sizes(1), sizes(2), 1, 1);
f_tmp = filt_package ([], f_template, f_template, shift, 's');

f_value = repmat(f_tmp.f1_s, [1 sizes(3)]);
f_i1 = repmat(f_tmp.i1_s, [1 sizes(3)]);
f_i2 = repmat(f_tmp.i2_s, [1 sizes(3)]);
idx_offset = [0:num_new_feature-1];
idx_offset = repmat(idx_offset, [sizes(1)*sizes(2)+1 1]);
idx_offset(end,:) = 0;
f_i3 = repmat(f_tmp.i3_s,[1 sizes(3)])+idx_offset;

f = setfield(f, ['f1_' tag], f_value);
f = setfield(f, ['f2_' tag], f_value);
f = setfield(f, ['i1_' tag], f_i1);
f = setfield(f, ['i2_' tag], f_i2);
f = setfield(f, ['i3_' tag], f_i3);
f = setfield(f, ['size_' tag], [sizes(1:2) sizes(3) sizes(3)]);
f = setfield(f, ['shift_' tag], shift);