test_idx = labels>200;
% prepare train images
images_test = mpfaces(:,:,:,test_idx);
[h,w,c,num_test] = size(images_test);
% prepare test views
views_test = views(:,test_idx);
theta_test = atand(views_test(1,:)./views_test(2,:)); theta_test = theta_test(:);
ids_theta = unique(theta_test);
dim3 = length(ids_theta);
vclasses_test = zeros(dim3,num_test);
for cc = 1:length(ids_theta),
    vclasses_test(cc,theta_test==ids_theta(cc)) = 1;
end
% prepare test class index
labels_test = labels(test_idx);
sessions_test = mpslabel(test_idx);
% prepare class index
ind_gallery = zeros(size(labels_test));
for cc = unique(labels_test)',
    session_cc = unique(sessions_test(labels_test==cc));
    ss = min(session_cc);
    ind = (labels_test==cc)&(sessions_test==ss)&(theta_test==0);
    ind_gallery = ind_gallery + ind;
end;
ind_gallery = ind_gallery > 0;
ind_probe = (~ind_gallery);
sum(ind_gallery)
feats_test = zeros(512,num_test,'single');
for m = 1:num_test
    images = images_test(:,:,:,m);
    images = reshape(single(images),[h,w,c,1]);
    images = permute(images,[2,1,3,4]);
    preds = caffe('forward', { images });
    feats_test(:,m) = squeeze(preds{1});
    if mod(m,100)==0, fprintf(sprintf('extract %d features\n', m)); end
end
feats_gallery = feats_test(:,ind_gallery);
feats_probe = feats_test(:,ind_probe);
labels_gallery = labels_test(ind_gallery);
labels_probe = labels_test(ind_probe);
theta_gallery = theta_test(ind_gallery);
theta_probe = theta_test(ind_probe);
dist = EuDist2(feats_gallery', feats_probe', 1);
[~,matches] = min(dist,[],1);
labels_pred = labels_gallery(matches);
%%
acc = [];
for t = [0,-15,15,-30,30,-45,45],
    labels_pred_sub = labels_pred(theta_probe==t);
    labels_probe_sub = labels_probe(theta_probe==t);
    acc = [acc; sum(labels_pred_sub==labels_probe_sub)/length(labels_probe_sub)];
end
acc

acc =

    0.5521
    0.5700
    0.5967
    0.1767
    0.3533
    0.0900
    0.0567