File size: 9,203 Bytes
2976de3 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 |
function [recall, precision, info] = vl_pr(labels, scores, varargin)
%VL_PR Precision-recall curve.
% [RECALL, PRECISION] = VL_PR(LABELS, SCORES) computes the
% precision-recall (PR) curve. LABELS are the ground truth labels,
% greather than zero for a positive sample and smaller than zero for
% a negative one. SCORES are the scores of the samples obtained from
% a classifier, where lager scores should correspond to positive
% samples.
%
% Samples are ranked by decreasing scores, starting from rank 1.
% PRECISION(K) and RECALL(K) are the precison and recall when
% samples of rank smaller or equal to K-1 are predicted to be
% positive and the remaining to be negative. So for example
% PRECISION(3) is the percentage of positive samples among the two
% samples with largest score. PRECISION(1) is the precision when no
% samples are predicted to be positive and is conventionally set to
% the value 1.
%
% Set to zero the lables of samples that should be ignored in the
% evaluation. Set to -INF the scores of samples which are not
% retrieved. If there are samples with -INF score, then the PR curve
% may have maximum recall smaller than 1, unless the INCLUDEINF
% option is used (see below). The options NUMNEGATIVES and
% NUMPOSITIVES can be used to add additional surrogate samples with
% -INF score (see below).
%
% [RECALL, PRECISION, INFO] = VL_PR(...) returns an additional
% structure INFO with the following fields:
%
% info.auc::
% The area under the precision-recall curve. If the INTERPOLATE
% option is set to FALSE, then trapezoidal interpolation is used
% to integrate the PR curve. If the INTERPOLATE option is set to
% TRUE, then the curve is piecewise constant and no other
% approximation is introduced in the calculation of the area. In
% the latter case, INFO.AUC is the same as INFO.AP.
%
% info.ap::
% Average precision as defined by TREC. This is the average of the
% precision observed each time a new positive sample is
% recalled. In this calculation, any sample with -INF score
% (unless INCLUDEINF is used) and any additional positive induced
% by NUMPOSITIVES has precision equal to zero. If the INTERPOLATE
% option is set to true, the AP is computed from the interpolated
% precision and the result is the same as INFO.AUC. Note that AP
% as defined by TREC normally does not use interpolation [1].
%
% info.ap_interp_11::
% 11-points interpolated average precision as defined by TREC.
% This is the average of the maximum precision for recall levels
% greather than 0.0, 0.1, 0.2, ..., 1.0. This measure was used in
% the PASCAL VOC challenge up to the 2008 edition.
%
% info.auc_pa08::
% Deprecated. It is the same of INFO.AP_INTERP_11.
%
% VL_PR(...) with no output arguments plots the PR curve in the
% current axis.
%
% VL_PR() accepts the following options:
%
% Interpolate:: false
% If set to true, use interpolated precision. The interpolated
% precision is defined as the maximum precision for a given recall
% level and onwards. Here it is implemented as the culumative
% maximum from low to high scores of the precision.
%
% NumPositives:: []
% NumNegatives:: []
% If set to a number, pretend that LABELS contains this may
% positive/negative labels. NUMPOSITIVES/NUMNEGATIVES cannot be
% smaller than the actual number of positive/negative entrires in
% LABELS. The additional positive/negative labels are appended to
% the end of the sequence, as if they had -INF scores (not
% retrieved). This is useful to evaluate large retrieval systems
% for which one stores ony a handful of top results for efficiency
% reasons.
%
% IncludeInf:: false
% If set to true, data with -INF score SCORES is included in the
% evaluation and the maximum recall is 1 even if -INF scores are
% present. This option does not include any additional positive or
% negative data introduced by specifying NUMPOSITIVES and
% NUMNEGATIVES.
%
% Stable:: false
% If set to true, RECALL and PRECISION are returned the same order
% of LABELS and SCORES rather than being sorted by decreasing
% score (increasing recall). Samples with -INF scores are assigned
% RECALL and PRECISION equal to NaN.
%
% NormalizePrior:: []
% If set to a scalar, reweights positive and negative labels so
% that the fraction of positive ones is equal to the specified
% value. This computes the normalised PR curves of [2]
%
% About the PR curve::
% This section uses the same symbols used in the documentation of
% the VL_ROC() function. In addition to those quantities, define:
%
% PRECISION(S) = TP(S) / (TP(S) + FP(S))
% RECALL(S) = TPR(S) = TP(S) / P
%
% The precision is the fraction of positivie predictions which are
% correct, and the recall is the fraction of positive labels that
% have been correctly classified (recalled). Notice that the recall
% is also equal to the true positive rate for the ROC curve (see
% VL_ROC()).
%
% REFERENCES:
% [1] C. D. Manning, P. Raghavan, and H. Schutze. An Introduction to
% Information Retrieval. Cambridge University Press, 2008.
% [2] D. Hoiem, Y. Chodpathumwan, and Q. Dai. Diagnosing error in
% object detectors. In Proc. ECCV, 2012.
%
% See also VL_ROC(), VL_HELP().
% Author: Andrea Vedaldi
% Copyright (C) 2007-12 Andrea Vedaldi and Brian Fulkerson.
% All rights reserved.
%
% This file is part of the VLFeat library and is made available under
% the terms of the BSD license (see the COPYING file).
% TP and FP are the vectors of true positie and false positve label
% counts for decreasing scores, P and N are the total number of
% positive and negative labels. Note that if certain options are used
% some labels may actually not be stored explicitly by LABELS, so P+N
% can be larger than the number of element of LABELS.
[tp, fp, p, n, perm, varargin] = vl_tpfp(labels, scores, varargin{:}) ;
opts.stable = false ;
opts.interpolate = false ;
opts.normalizePrior = [] ;
opts = vl_argparse(opts,varargin) ;
% compute precision and recall
small = 1e-10 ;
recall = tp / max(p, small) ;
if isempty(opts.normalizePrior)
precision = max(tp, small) ./ max(tp + fp, small) ;
else
a = opts.normalizePrior ;
precision = max(tp * a/max(p,small), small) ./ ...
max(tp * a/max(p,small) + fp * (1-a)/max(n,small), small) ;
end
% interpolate precision if needed
if opts.interpolate
precision = fliplr(vl_cummax(fliplr(precision))) ;
end
% --------------------------------------------------------------------
% Additional info
% --------------------------------------------------------------------
if nargout > 2 || nargout == 0
% area under the curve using trapezoid interpolation
if ~opts.interpolate
if numel(precision) > 1
info.auc = 0.5 * sum((precision(1:end-1) + precision(2:end)) .* diff(recall)) ;
else
info.auc = 0 ;
end
end
% average precision (for each recalled positive sample)
sel = find(diff(recall)) + 1 ;
info.ap = sum(precision(sel)) / p ;
if opts.interpolate
info.auc = info.ap ;
end
% TREC 11 points average interpolated precision
info.ap_interp_11 = 0.0 ;
for rc = linspace(0,1,11)
pr = max([0, precision(recall >= rc)]) ;
info.ap_interp_11 = info.ap_interp_11 + pr / 11 ;
end
% legacy definition
info.auc_pa08 = info.ap_interp_11 ;
end
% --------------------------------------------------------------------
% Plot
% --------------------------------------------------------------------
if nargout == 0
cla ; hold on ;
plot(recall,precision,'linewidth',2) ;
if isempty(opts.normalizePrior)
randomPrecision = p / (p + n) ;
else
randomPrecision = opts.normalizePrior ;
end
spline([0 1], [1 1] * randomPrecision, 'r--', 'linewidth', 2) ;
axis square ; grid on ;
xlim([0 1]) ; xlabel('recall') ;
ylim([0 1]) ; ylabel('precision') ;
title(sprintf('PR (AUC: %.2f%%, AP: %.2f%%, AP11: %.2f%%)', ...
info.auc * 100, ...
info.ap * 100, ...
info.ap_interp_11 * 100)) ;
if opts.interpolate
legend('PR interp.', 'PR rand.', 'Location', 'SouthEast') ;
else
legend('PR', 'PR rand.', 'Location', 'SouthEast') ;
end
clear recall precision info ;
end
% --------------------------------------------------------------------
% Stable output
% --------------------------------------------------------------------
if opts.stable
precision(1) = [] ;
recall(1) = [] ;
precision_ = precision ;
recall_ = recall ;
precision = NaN(size(precision)) ;
recall = NaN(size(recall)) ;
precision(perm) = precision_ ;
recall(perm) = recall_ ;
end
% --------------------------------------------------------------------
function h = spline(x,y,spec,varargin)
% --------------------------------------------------------------------
prop = vl_linespec2prop(spec) ;
h = line(x,y,prop{:},varargin{:}) ;
|