Description of slverifyroc_blks

Home > sltoolbox > perfeval > slverifyroc_blks.m

slverifyroc_blks

PURPOSE

SLVERIFYROC_BLKS Computes the verification ROC for blockwise score matrix

SYNOPSIS

function [thrs, fars, frrs] = slverifyroc_blks(scores, blocks, labels1, labels2, op, npts)

DESCRIPTION

SLVERIFYROC_BLKS Computes the verification ROC for blockwise score matrix

 $ Syntax $
   - [thrs, fars, frrs] = slverifyroc_blks(scores, blocks, labels1, labels2, op)
   - [thrs, fars, frrs] = slverifyroc_blks(scores, blocks, labels1, labels2, op, npts)
 
 $ Arguments $
   - scores:       the cell array of array filenames of the scores
   - blocks:       the division blocks
   - labels1:      the labels of referred samples
   - labels2:      the labels of query samples
   - op:           the option stating the attributes of the scores
   - npts:         the number of threshold points (default = 500)
   - thrs:         the sampled thresholds
   - fars:         the false accept rates at the sampled thresholds
   - frrs:         the false reject rates at the sampled thresholds

 $ Remarks $
   - This function is an extension of slverifyroc to support blockwise
     scores in large scale experiments. However, the implementation is
     fundamentally changed.

 $ History $
   - Created by Dahua Lin, on Aug 8th, 2006

CROSS-REFERENCE INFORMATION

This function calls:

• slreadarray SLREADARRAY Reads an array from an array file
• slhistroc SLHISTROC Computes the ROC curve from value histogram
• raise_lackinput RAISE_LACKINPUT Raises an error indicating lack of input argument

This function is called by:

SUBFUNCTIONS

• function [maxv, minv] = collect_maxmin(S)
• function H = collect_scorehist(S, l1, l2, thrs)

SOURCE CODE

0001 function [thrs, fars, frrs] = slverifyroc_blks(scores, blocks, labels1, labels2, op, npts)
0002 %SLVERIFYROC_BLKS Computes the verification ROC for blockwise score matrix
0003 %
0004 % $ Syntax $
0005 %   - [thrs, fars, frrs] = slverifyroc_blks(scores, blocks, labels1, labels2, op)
0006 %   - [thrs, fars, frrs] = slverifyroc_blks(scores, blocks, labels1, labels2, op, npts)
0007 %
0008 % $ Arguments $
0009 %   - scores:       the cell array of array filenames of the scores
0010 %   - blocks:       the division blocks
0011 %   - labels1:      the labels of referred samples
0012 %   - labels2:      the labels of query samples
0013 %   - op:           the option stating the attributes of the scores
0014 %   - npts:         the number of threshold points (default = 500)
0015 %   - thrs:         the sampled thresholds
0016 %   - fars:         the false accept rates at the sampled thresholds
0017 %   - frrs:         the false reject rates at the sampled thresholds
0018 %
0019 % $ Remarks $
0020 %   - This function is an extension of slverifyroc to support blockwise
0021 %     scores in large scale experiments. However, the implementation is
0022 %     fundamentally changed.
0023 %
0024 % $ History $
0025 %   - Created by Dahua Lin, on Aug 8th, 2006
0026 %
0027 
0028 %% parse and verify input arguments
0029 
0030 if nargin < 5
0031     raise_lackinput('slverifyroc_blks', 5);
0032 end
0033 if isempty(labels2)
0034     labels2 = labels1;
0035 end
0036 if ~iscell(scores)
0037     error('sltoolbox:invalidargs', ...
0038         'The scores should be a cell array of filenames');
0039 end
0040 if nargin < 6 || isempty(npts)
0041     npts = 500;
0042 end
0043 nblks = numel(scores);
0044 if ~isequal(size(scores), size(blocks))
0045     error('The sizes of scores and blocks are inconsistent');
0046 end
0047 
0048 
0049 %% Collect Histogram
0050 
0051 % collect min, max value
0052 maxv = -inf;
0053 minv = inf;
0054 for k = 1 : nblks
0055     [curmaxv, curminv] = collect_maxmin(slreadarray(scores{k}));
0056     if curmaxv > maxv
0057         maxv = curmaxv;
0058     end
0059     if curminv < minv
0060         minv = curminv;
0061     end
0062 end
0063 
0064 if minv >= maxv
0065     error('sltoolbox:valuerror', ...
0066         'The minv should be less than maxv');
0067 end
0068 
0069 % determine thresholds
0070 thrs = linspace(minv, maxv, npts)';
0071 
0072 % collect histograms
0073 
0074 H = zeros(npts, 2);
0075 for k = 1 : nblks
0076     cb = blocks{k};
0077     l1 = labels1(cb(1,1):cb(2,1));
0078     l2 = labels2(cb(1,2):cb(2,2));
0079     curscores = slreadarray(scores{k});    
0080     curH = collect_scorehist(curscores, l1, l2, thrs);
0081     
0082     H = H + curH;
0083 end
0084 
0085 
0086 %% Compute ROC
0087 
0088 hist_a = H(:, 1);
0089 hist_r = H(:, 2);
0090 
0091 [thrs, fars, frrs] = slhistroc(hist_a, hist_r, thrs, op);
0092 
0093 
0094 %% The internal functions
0095 
0096 function [maxv, minv] = collect_maxmin(S)
0097 
0098 S = S(:);
0099 maxv = max(S);
0100 minv = min(S);
0101 
0102 
0103 function H = collect_scorehist(S, l1, l2, thrs)
0104 % H is an nbins x 2 matrix stored as [hist_a, hist_r]
0105 
0106 m = length(l1);
0107 n = length(l2);
0108 if ~isequal(size(S), [m, n])
0109     error('sltoolbox:sizmismatch', ...
0110         'The sizes of labels and score matrix are mismatch');
0111 end
0112 
0113 l1 = l1(:);
0114 l2 = l2(:)';
0115 L1 = l1(:, ones(1, n));
0116 L2 = l2(ones(m, 1), :);
0117 signals = (L1 == L2);
0118 clear L1 L2;
0119 
0120 scores_a = S(signals);
0121 scores_r = S(~signals);
0122 
0123 if ~isempty(scores_a)
0124     hist_a = histc(scores_a, thrs);
0125 else
0126     hist_a = zeros(length(thrs), 1);
0127 end
0128 
0129 if ~isempty(scores_r)
0130     hist_r = histc(scores_r, thrs);
0131 else
0132     hist_r = zeros(length(thrs), 1);
0133 end
0134 
0135 H = [hist_a, hist_r];
0136 
0137

---

Generated on Wed 20-Sep-2006 12:43:11 by m2html © 2003