generateTemporalPairwise.m

function probality = generateTemporalPairwise( sequence, densematch, model, normalize_ratio)
addpath(genpath('/BS/pedestrian-detection-tracking/work/project/3rd/liblinear-1.94/'));
addpath /home/tang/Projects/multicut_tracking/tracking-multicut-git/utils
% In this function we generate edges and cost of edges detections in
% neighboring frames
% sequence - the structure contains image list and detection list
% densematch: in densematch.mat
% model: obtained by loading the pairwise model
% normalize_ratio: obtained by loading the pairwise model

% load data from eldar
img_list = {sequence.image};
detection_list = {sequence.detections};
num_frames = length(img_list);
assert (length(detection_list) == num_frames);
img_size = size(imread(img_list{1}));

% hard code part size
box_size = 40;
% convert detection coordinates to bounding boxes
detection_all = cell(num_frames,1);
for i = 1:num_frames
    det = load(detection_list{i});det = det.detection_coordinates;
    box = [det(:,1)-box_size/2, det(:,2)-box_size/2,det(:,1)+box_size/2, det(:,2)+box_size/2];
    detection_all{i} = box;
end

% compute pairwise probality between detections in the neighboring frames
probality = cell(num_frames);
for i = 1:num_frames -1
    fprintf('Compute pairwise cost for frame:  %d \n', i);
    
    det_1 = detection_all{i};
    det_2 = detection_all{i+1};
    cur_dm = densematch{i,i+1};
    cur_prob = compute_pairwise_cost_block_densematch(det_1,det_2,model,normalize_ratio,cur_dm,img_size);
    probality{i,i+1} = cur_prob;
end
end
function edges_prob = compute_pairwise_cost_block_densematch(b1,b2,model,normalize_ratio,cur_dm,img_size)

num_det_1 = size(b1,1);
num_det_2 = size(b2,1);

b1(b1(:,1)<=0,1) = 1;
b1(b1(:,2)<=0,2) = 1;
b1(b1(:,3)==0,3) = 1;b1(b1(:,3)> img_size(2),3) = img_size(2);
b1(b1(:,4)==0,4) = 1;b1(b1(:,4)> img_size(1),4) = img_size(1);

b2(b2(:,1)<=0,1) = 1;
b2(b2(:,2)<=0,2) = 1;
b2(b2(:,3)==0,3) = 1;b2(b2(:,3)> img_size(2),3) = img_size(2);
b2(b2(:,4)==0,4) = 1;b2(b2(:,4)> img_size(1),4) = img_size(1);

f_dm = zeros(num_det_2,num_det_1);
parfor i = 1 : num_det_1
    d1 = b1(i,:); rowidx1 = ceil(d1(2)):floor(d1(4));colidx1 = ceil(d1(1)):floor(d1(3));
    rowsub1 = repmat(rowidx1',1,length(colidx1));
    colsub1 = repmat(colidx1,length(rowidx1),1);
    d1_ind = sub2ind(img_size(1:2), rowsub1(:), colsub1(:));
    [~,ia_1,~] = intersect(cur_dm(:,1),d1_ind);
    for j = 1:num_det_2
        d2 = b2(j,:); rowidx2 = ceil(d2(2)):floor(d2(4));colidx2 = ceil(d2(1)):floor(d2(3));
        rowsub2 = repmat(rowidx2',1,length(colidx2));
        colsub2 = repmat(colidx2,length(rowidx2),1);
        d2_ind = sub2ind(img_size(1:2), rowsub2(:), colsub2(:));
        [~,ia_2,~] = intersect(cur_dm(:,2),d2_ind);
        
        i_box = intersect(ia_1,ia_2);
        u_box = union(ia_1,ia_2);
        if isempty(u_box)
            f_dm(j,i) = 0
        else
            f_dm(j,i) = length(i_box)/length(u_box);
        end
    end
end

f = [f_dm(:)];

[exp_num, dim] = size(f);

f_quad = zeros(exp_num, (dim-1)*dim/2);
dim_count = 1;
for m = 1:dim
    for n = 1:dim
        f_quad(:,dim_count) = f(:,m).*f(:,n);
        dim_count = dim_count + 1;
    end
end
f = [f f_quad];

%%
f =  normalize_feature(f, normalize_ratio);

f = sparse(double(f));
label = ones(size(f,1),1);
[~, ~,pro_estimate] = predict(label, f, model,  '-b 1 -q');

edges_prob = reshape(pro_estimate(:,1),num_det_2,num_det_1);
end