filt_get_c1.m

Standard model object recognition matlab code

function f = filt_get_C1

% FUNCTION f = filt_get_C1
%
% Here, we define which scales and spatial positions to pool, so
% that C1 units are translation and scale invariant complex cells
% in V1.  Many parameters (eg. number of scales/orientations) are
% carried over from S1 filter parameters.

f = [];
num_scale = 8; % there are 8 scales in S1.
for i = 1:num_scale
  scale_tag = ['s' num2str(i)];
  [f1, f2, shift] = get_C1_filters(i);
  f = filt_package (f, f1, f2, shift, scale_tag);
end

function [f1, f2, shift] = get_C1_filters (which_scale)

% FUNCTION [f1, f2, shift] = get_C1_filters (which_scale)
% 
% Defines C1 filter parameters, so make sure the parameters match
% with the definitions from S1 filters.
% In the old, HMAX parameters (Nature Neurosci. 1999), 
%    pooling_ranges = [4 6 9 12];
%    shift          = ceil(pooling_range/2);

% shift values are hard-coded to make the receptive field sizes of
% C1 units (S1 filter size plus pooling range) divided by the shift
% values are constant (eg. same overlap factors or sampling across
% different scales).  This is important for scale invariance. 

% Parameter with T. Serre.
%shifts = [3 5 7 8 10 12 13 15];

% However, below is more correct.
%shifts = round([16 22 28 34 40 46 52 59]/3);
%shifts = [3 4 5 7 8 9 10 12];
%shift  = shifts(which_scale);

num_orientation = 4;

[a,b,c] = init_filter_def(2); % Grab C1 definitions.
pooling_ranges = b(1,:);
num_old_feats  = b(3,:);
num_new_feats  = b(4,:);
shifts         = c;

pooling_range  = pooling_ranges(which_scale);
num_old_feat   = num_old_feats(which_scale);
num_new_feat   = num_new_feats(which_scale);
shift          = c(which_scale);

max_fs       = pooling_range; % max filter size

f = zeros(max_fs,max_fs,num_old_feat,num_new_feat);

switch which_scale
  % Based on S1 filter parameters
  case 8
    num_scale = 3;
  otherwise
    num_scale = 2;
end
for i = 1:num_orientation
  f(:,:,[1:num_scale]+(i-1)*num_scale,i) = 1;
end

f1 = f;
f2 = f;