ExtractFeatures.m

close all
clear all
clc


%% set images path and ground-truth path
imagesPath = 'part_A/train_data/images/';
groundPath = 'part_A/train_data/ground-truth';
n = 182; % images number
features = cell(1, n);
counts = cell(1, n);

winSize = 100;
winStep = winSize - 1;
xStep = 50;
yStep = 50;


%% init pre-trained resnet-152 model 
net = dagnn.DagNN.loadobj(load('imagenet-resnet-152-dag.mat'));
net.mode = 'test';
net.conserveMemory = 0;

%% 
for i = 1 : n
    disp(i)
    im = imread([imagesPath 'IMG_' num2str(i) '.jpg']);
    load([groundPath 'GT_IMG_' num2str(i) '.mat']);
    [height, width, channel] = size(im);
    
    newHeight = round(height/50) * 50;
    newWidth = round(width/50) * 50;

    location = image_info{1}.location;
    location(:, 1) = location(:, 1) / width * newWidth;
    location(:, 2) = location(:, 2) / height * newHeight;

    im = imresize(im, [newHeight, newWidth]);
    if channel == 1
        tmp = zeros(newHeight, newWidth, 3);
        tmp(:, :, 1) = im;
        tmp(:, :, 2) = im;
        tmp(:, :, 3) = im;
        im = tmp;
    end
    
    
    y = 1;
    row = 1;
    patchFeature = zeros(newHeight / 50 - 1, newWidth/50 - 1, 1000);
    patchCount = zeros(newHeight / 50 - 1, newWidth/50 - 1);
    while(y + winStep  <= newHeight)
        x = 1;
        column = 1;
        while(x + winStep <= newWidth)
            img = im(y:y + winStep, x: x + winStep, :);% get image patch
            img = single(img);
            
            im_ = imresize(img, net.meta.normalization.imageSize(1:2));
            im_ = im_ - net.meta.normalization.averageImage ;
            
            net.eval({'data', im_});
            patchFeature(row ,column ,:) = reshape(net.vars(net.getVarIndex('fc1000')).value, 1, 1000);
            
            index =  (location(:, 1) > x - 0.5) & (location(:, 1) < x + winStep + 0.5 ) & (location(:, 2) > y - 0.5) & (location(:, 2) < y + winStep + 0.5);
            patchCount(row, column) = sum(index);
            
            x = x + xStep;
            column = column + 1;
        end
        y = y + yStep;
        row = row + 1;
    end
    
    features{i} = patchFeature;
    counts{i} = patchCount;
end

save train_resnet_A.mat features counts