function showDet(detDir, className, aspectName, suppThresh) % SHOWDET % % Author:: Andrea Vedaldi % AUTORIGHTS % Copyright (C) 2008-09 Andrea Vedaldi % % This file is part of the VGG MKL Class and VGG MKL Det code packages, % available in the terms of the GNU General Public License version 2. if nargin < 3, aspectName = '.*' ; end if nargin < 4, suppThresh = NaN ; end if ischar(suppThresh), suppThresh = sscanf(suppThresh, '%f'); end conf = do_conf_fun() ; fprintf('Class name: %s\n', className) ; fprintf('Aspect name: %s\n', aspectName) ; if ~isnan(suppThresh), fprintf('Suppress threshold: %.2f\n', suppThresh) ; end ; % -------------------------------------------------------------------- % Load databases % -------------------------------------------------------------------- fprintf('Loading databases.\n') ; imdb = getImageDb(conf.imageDbPath) ; roidb = getRoiDb(conf.gtRoiDbPath) ; imageIndex = [imdb.images.id] ; imageNames = char({imdb.images.name}) ; % -------------------------------------------------------------------- % Load candidates % -------------------------------------------------------------------- fprintf('Reading candidates for ''%s''.\n', detDir) ; candList = cacheGet(mfilename, ... {detDir, className, aspectName, suppThresh}, {}) ; if isempty(candList) detDirList = dir(detDir) ; if isempty(detDirList) error(sprintf('The directory ''%s'' is empty', detDir)) ; end candList = {} ; n = 1 ; for detDirEntry = detDirList' extract = regexp(detDirEntry.name, '^(?[0-9]\w+).mat$', 'names') ; if isempty([extract.imageName]), continue ; end candList{n} = load(fullfile(detDir, detDirEntry.name)) ; candList{n}.imageName = extract.imageName ; if ~isnan(suppThresh) sel = suppressBoxes(candList{n}.box, suppThresh, 5) ; % sel = sel(1:5) ; candList{n}.box = candList{n}.box(:,sel) ; candList{n}.score = candList{n}.score(:,sel) ; end n = n + 1 ; end candList = cat(2, candList{:}) ; sourcedImageNames = char({candList.imageName}) ; [ok, ii] = ismember(sourcedImageNames, imageNames, 'rows') ; tmp = {imdb.images(ii).id} ; [candList.imageId] = deal(tmp{:}) ; % remove stuff which is not in test... sel = find([imdb.images(binsearch(imageIndex, [candList.imageId])).set] == imdb.sets.TEST07) ; candList = candList(sel) ; if ~all(ok), error('Some of the image could not be found in DB') ; end cacheStore(mfilename, candList, ... {detDir, className, aspectName, suppThresh}, {}, 1) ; end % -------------------------------------------------------------------- % Match candidates to GT % -------------------------------------------------------------------- fprintf('Comparing to GT.\n') ; roiList = cacheGet([mfilename '_roi'], ... {detDir, className, aspectName, suppThresh}, {}) ; if isempty(roiList) roiList = {} ; for ci = 1:length(candList) match = matchToGt(ci) ; selMiss = find(isnan(match.gtBoxToDet)) ; missBoxes = roidb.rois.box(:, sel(selMiss)) ; missScores = -inf * ones(1, size(missBoxes,2)) ; missFlags = ones(1, size(missBoxes,2)) ; roiList{ci}.candSel = ci * ones(1, length(candList(ci).score) + size(missBoxes,2)) ; roiList{ci}.scores = [candList(ci).score missScores] ; roiList{ci}.boxes = [candList(ci).box missBoxes] ; roiList{ci}.flags = [match.detBoxFlags missFlags] ; end roiList = aosToSoa(cat(2, roiList{:})) ; cacheStore([mfilename '_roi'], roiList, ... {detDir, className, aspectName, suppThresh}, {}, 1) ; end % -------------------------------------------------------------------- % Display % -------------------------------------------------------------------- fprintf('Instantiating GUI.\n') ; % sort candidates by best match score bestScores = arrayfun(@(x) max([x.score -inf]), candList) ; worstScores = arrayfun(@(x) min([x.score +inf]), candList) ; bestScore = max(bestScores) ; worstScore = min(worstScores) ; [dorp, candOrder] = sort(bestScores,'descend') ; [drop, roiOrder] = sort(roiList.scores, 'descend') ; recall = cumsum(roiList.flags(roiOrder) == +1) / sum(roiList.flags(roiOrder) == +1) ; precision = cumsum(roiList.flags(roiOrder) == +1) ./ ((cumsum(roiList.flags(roiOrder) ~= 0) + eps)) ; stop = max(find(roiList.scores(roiOrder) > -inf)) ; %fh = figure(101) ; clf ; fh = figure ; clf ; %subplot(1,2,1) ; plot(roiList.scores(roiOrder)) ; title('ROI scores') ; %subplot(1,2,2) ; plot(recall, precision) ; grid on ; axis equal ; xlim([0,1]) ; ylim([0,1]) ; [a,a08] = auc([0 recall(1:stop)], [1 precision(1:stop)]) ; title(sprintf('precision-recall %.2f %% (%.2f %%)', a*100, a08*100)) ; drawnow ; candLabels = arrayfun(@(x) sprintf('%012.0f', candList(x).imageId), ... candOrder, ... 'UniformOutput', false) ; roiLabels = arrayfun(@(x) sprintf('%2.3f %3d [%6.2f %6.2f]', ... roiList.scores(roiOrder(x)), roiList.flags(roiOrder(x)), ... 100*recall(x), 100*precision(x)), ... 1:length(roiOrder), ... 'UniformOutput', false) ; fh = figure ; clf ; set(gcf,'units', 'normalized') ; candSel = uicontrol(fh, 'style', 'listbox', ... 'String', candLabels, ... 'Value', 1, ... 'Units', 'normalized', 'Position', [0 0 .10 1], ... 'Callback', @displayCand) ; roiSel = uicontrol(fh, 'style', 'listbox', ... 'String', roiLabels, ... 'Value', 1, ... 'Units', 'normalized', 'Position', [.10 0 .15 1], ... 'Callback', @displayRoi) ; infoH = axes('position', [.25 0 .25 1]) ; imgH = axes('position', [.5 0 .5 .5]) ; gtH = axes('position', [.5 .5 .5 .5]) ; display(1) ; function [match,gtRois] = matchToGt(ci) % ~~~~~~~~~~~~~ cands = candList(ci) ; ii = binsearch(imageIndex, cands.imageId) ; imageId = imdb.images(ii).id ; % find instances of this class sel = findRois(roidb, ... 'imageId', imageId, ... 'jitter', 0, ... 'class', roidb.classes.(upper(className))) ; % and specifically of this aspect selAspect = findRois(roidb, ... 'subset', sel, ... 'aspect', aspectName) ; % difficult are also other aspects gtRois = soaSubsRef(roidb.rois, sel) ; gtRois.difficult(~ismember(sel, selAspect)) = 1 ; match = evalDetections(gtRois.box, ... gtRois.difficult, ... candList(ci).box, ... candList(ci).score) ; end function displayCand(varargin) % ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ci = candOrder(get(candSel, 'value')) ; display(ci) end function displayRoi(varargin) % ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ri = roiOrder(get(roiSel, 'value')) ; display(roiList.candSel(ri)) ; axes(imgH) ; plotroi(roiList.boxes(:,ri), 'b', 'linewidth', 1) ; axes(gtH) ; plotroi(roiList.boxes(:,ri), 'b', 'linewidth', 1) ; end function display(ci) % ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ cands = candList(ci) ; ii = binsearch(imageIndex, cands.imageId) ; imagePath = fullfile(imdb.dir, [imdb.images(ii).name '.jpg']) ; imageId = imdb.images(ii).id ; imageSize = imdb.images(ii).size ; imageName = imdb.images(ii).name ; imageSet = decodeEnum(imdb.sets, imdb.images(ii).set) ; im = imread(imagePath) ; [match, gtRois] = matchToGt(ci) ; str = '' ; str = [str sprintf('num cands:\t%d', length(cands.box))] ; str = [str sprintf('\nmax score:\t%g', max(cands.score))] ; str = [str sprintf('\nmin score:\t%g', min(cands.score))] ; str = [str sprintf('\n')] ; str = [str sprintf('\nimg id:\t%d', imageId)] ; str = [str sprintf('\nimg name:\t%s', imageName)] ; str = [str sprintf('\nimg set:\t%s', imageSet)] ; str = [str sprintf('\nimg geom:\t%dx%d', imageSize(1), imageSize(2))] ; str = [str sprintf('\n')] ; cla(gtH) ; axes(gtH) ; image(im,'parent',gtH) ; hold(gtH,'on') ; axis equal off ; for bi = size(cands.box,2):-1:1 ; switch match.detBoxFlags(bi) case +1, cl = 'g' ; case -1, cl = 'r' ; case 0, cl = 'y' ; end plotroi(cands.box(:,bi), 'linewidth', 2, 'color', cl) ; end for gti = 1:length(gtRois.id) if gtRois.difficult(gti) plotroi(gtRois.box(:, gti),'w--', 'linewidth', 3, ... 'label', sprintf('%d', gti)) ; else plotroi(gtRois.box(:, gti), 'w-', 'linewidth', 3, ... 'label', sprintf('%d', gti)) ; end str = [str sprintf('\n%d [%012.0f]',gti, gtRois.id(gti))] ; str = [str sprintf('\n aspect:\t%s', decodeEnum(roidb.aspects, gtRois.aspect(gti)))] ; str = [str sprintf('\n difficult:\t%d',gtRois.difficult(gti))] ; end cla(imgH) ; axes(imgH) ; image(im,'parent',imgH) ; hold(imgH,'on') ; axis equal off ; cl = colorizeScore(cands.score, bestScore, worstScore) ; for bi = size(cands.box,2):-1:1 ; plotroi(cands.box(:,bi), 'linewidth', 2, 'color', cl(bi,:)) ; end cla(infoH) ; axes(infoH) ; text(0,0, str, ... 'Backgroundcolor', 'w', ... 'VerticalAlign', 'top', ... 'Interpreter','none') ; set(gca,'ydir', 'reverse') ; axis off ; drawnow ; end % display end % main % -------------------------------------------------------------------- function cl = colorizeScore(score, bestScore, worstScore) % -------------------------------------------------------------------- selp = find(score >= 0) ; seln = find(score < 0) ; lamp = score(selp) / bestScore ; lamn = score(seln) / worstScore ; cl = zeros(length(score), 3) ; cl(selp, :) = lamp(:) * [0 1 0] + (1 - lamp(:)) * [.5 .5 .5] ; cl(seln, :) = lamn(:) * [1 0 0] + (1 - lamn(:)) * [.5 .5 .5] ; end % colorizeScore