eval_dataset.m

%% load test set
load ../FLIC/examples.mat
examples = flip_backwards_facing_groundtruth(examples);
testex = examples([examples.istest]);
ntest = length(testex);
model = loadvar('MODEC-model.mat','mdls');

%% Run MODEC over every test example, saving wrist, elbow and shoulder locations.
% Takes 15-20 minutes.  Parallelizing is trivial.
pred_coords = nan(2,6,ntest);
for i=1:ntest
    tstart = clock;
    img = imread(fullfile('../FLIC/images/',testex(i).filepath));
    pred_i = run_modec(model, img, testex(i).torsobox);
    pred_coords(:,:,i) = pred_i.coords;
    fprintf('example %d / %d took %s\n',i,ntest,sec2timestr(etime(clock,tstart)));
end
%% evaluate
gt_coords = cat(3,testex.coords);

% error measure described in paper:
scale_by_parts = {'lsho','rhip'};
elbow_err = score_predictions(pred_coords, gt_coords, {'lelb','relb'}, scale_by_parts);
wrist_err = score_predictions(pred_coords, gt_coords, {'lwri','rwri'}, scale_by_parts);

% curves should match "MODEC+cascade" curves in Figure 4 of the paper.
clf
range = 1:20;
[~, elbow_auc] = accuracyCurve(elbow_err(:),range,'b-','linewidth',3)
hold on
[~, wrist_auc] = accuracyCurve(wrist_err(:),range,'g-','linewidth',3)
axis square, grid on
axis([range([1 end]) 1 100])
legend({'MODEC+cascade elbow acc', 'MODEC+cascade wrist acc'})
%% validate to ensure you get official results:
assert(abs(elbow_auc - 47.33993) < 1e-5)
assert(abs(wrist_auc - 35.36703) < 1e-5)