Add files via upload

bag_of_hogs.m

@@ -0,0 +1,120 @@
+function [image_features] = bag_of_hogs(vocabulary,type_of_dataset) 
+
+
+%Step size for SIFT detection
+step_size=4;
+vocab_size = size(vocabulary, 1);
+
+if(type_of_dataset)
+    plane_source = 'D:\Mandeep\Summer\BTP\Tracking\Hog\tight_dataset\';
+    non_plane_source = 'D:\Mandeep\Summer\BTP\Tracking\Hog\Negatives2\';
+    fprintf('Will find features for training datasets\n');
+    total_images_positive = 1000;
+    total_images_negative = 1000;
+    random_numbers=randi([1 2000],1,total_images_negative);    
+%     image_features = zeros(total_images, vocab_size);
+else
+    plane_source = 'D:\Mandeep\Summer\BTP\Tracking\Hog\positives_test_dataset\';
+    non_plane_source = 'D:\Mandeep\Summer\BTP\Tracking\Hog\negatives_test_dataset\';    
+    fprintf('Will find features for testing datasets \n');
+    total_images_positive = 5;
+    total_images_negative = 5;
+    random_numbers=randi([1 20],1,total_images_negative);
+%     image_features = zeros(total_images, vocab_size);
+end   
+%generating histograms of 2000 bins , where value of bin tells most no. of
+%matches with different cluster centres
+
+
+%for plane dataset
+
+counter = 1;
+image_counter = 1;
+fprintf('Computing Bag of words : Positives\n');
+while(image_counter <= total_images_positive)
+        filename = strcat(plane_source,num2str(counter),'.jpg');
+        if exist(filename,'file')
+            if (mod(image_counter,100)==0)
+                fprintf('      ..Processed %d Images\n',image_counter);
+                %fprintf(' Size of hog features : \n');
+                %size(hog_features)                
+            end
+            try
+            img = im2single(imread(filename));
+            catch
+                fprintf('Probelm in file %d',counter);
+                counter = counter + 1;
+                continue
+            end
+            %gives 128 x num_of_features with one descriptor per column
+            hog = vl_hog(img,step_size);
+            [x,y,z] = size(hog);
+            hog_features = (reshape(hog,[x*y,z]))';
+            %to reduce precision to 32
+            features = single(hog_features);
+            size(features)
+            [indices, distances] = knnsearch(vocabulary, features');
+
+            %generating histogram and then normalizing it
+            imhist=histc(indices, 1:vocab_size);
+            imhist_norm=imhist./numel(imhist);
+            image_features(image_counter,:) = imhist_norm';      
+            image_counter = image_counter + 1;
+            counter = counter + 1;
+        else
+            counter = counter + 1;
+        end    
+end
+
+fprintf('Computing Bag of words : Negatives\n');
+counter2 = 1;
+
+% for non_vehicle_dataset
+image_counter = 1;
+while(image_counter <= total_images_negative)    
+        filename = strcat(non_plane_source,num2str(random_numbers(counter2)),'.jpg');
+        if exist(filename,'file')
+            if (mod(image_counter,100)==0)
+                fprintf('      ..Processed %d Images\n',image_counter);
+                %fprintf(' Size of hog features : \n');
+                %size(hog_features)
+            end
+            try
+            img = im2single(imread(filename));
+%             img = imresize(img,0.2);
+            catch
+                fprintf('Probelm in file %d',counter2);
+                counter2 = counter2 + 1;
+                continue
+            end
+            %fprintf('Processed %d Images\n',image_counter);
+            hog = vl_hog(img,step_size);
+            [x,y,z] = size(hog);
+            hog_features = (reshape(hog,[x*y,z]))';
+            %to reduce precision to 32
+            features = single(hog_features);
+
+            %Find the nearest cluster center in vocabulary for each local feature
+            %in the image based on the Euclidean distance
+            %[indices, distances] = KNNSEARCH(vocabulary,features) returns a vector
+            %   distances containing the distances between each row of features and its
+            %   closest point in vocabulary. Each row in 'indices' contains the index of
+            %   the nearest neighbor in vocabulary for the corresponding row in features.
+            [indices, distances] = knnsearch(vocabulary, features');
+
+            %generating histogram and then normalizing it
+            hist=histc(indices, 1:vocab_size);
+            hist_norm=hist./numel(hist);
+            image_features2(image_counter,:) = hist_norm';
+
+            image_counter = image_counter + 1;
+            counter2 = counter2 + 1;
+        else
+            counter2 = counter2 + 1;
+        end
+end
+        fprintf('Size image features : %d\n', size(image_features,1));
+        fprintf('Size image features2 : %d\n', size(image_features2,1));
+        image_features = [image_features; image_features2];
+        fprintf('Size image features again: %d\n', size(image_features,1));
+end        

bag_of_sifts.m

@@ -0,0 +1,123 @@
+function [image_features] = bag_of_sifts(vocabulary,type_of_dataset) 
+
+
+%Step size for SIFT detection
+step_size=4;
+size(vocabulary);
+vocab_size = size(vocabulary, 1);
+
+if(type_of_dataset)
+    plane_source = 'D:\Mandeep\Summer\BTP\Tracking\Hog\positives_training_data\';
+    non_plane_source = 'D:\Mandeep\Summer\BTP\Tracking\Hog\negatives_training_data\';
+    fprintf('Will find features for training datasets\n');
+    total_images_positive = 2000;
+    total_images_negative = 5107; %3000
+    random_numbers=randi([1 5000],1,total_images_negative);    
+%     image_features = zeros(total_images, vocab_size);
+else
+    plane_source = 'D:\Mandeep\Summer\BTP\Tracking\Hog\positives_test_dataset\';
+    non_plane_source = 'D:\Mandeep\Summer\BTP\Tracking\Hog\negatives_test_dataset\';    
+    fprintf('Will find features for testing datasets \n');
+    total_images_positive = 5;
+    total_images_negative = 5;
+    random_numbers=randi([1 20],1,total_images_negative);
+%     image_features = zeros(total_images, vocab_size);
+end   
+%generating histograms of 2000 bins , where value of bin tells most no. of
+%matches with different cluster centres
+
+
+%for plane dataset
+
+counter = 1;
+image_counter = 1;
+fprintf('Computing Bag of words : Positives\n');
+while(image_counter <= total_images_positive)
+        filename = strcat(plane_source,num2str(counter),'.jpg');
+        if exist(filename,'file')
+            if (mod(image_counter,100)==0)
+                fprintf('      ..Processed %d Images\n',image_counter);
+                %fprintf(' Size of hog features : \n');
+                %size(hog_features)                
+            end
+            try
+            img = im2single(rgb2gray(imread(filename)));
+            catch
+                fprintf('Probelm in file %d',counter);
+                counter = counter + 1;
+                continue
+            end
+            %gives 128 x num_of_features with one descriptor per column
+            %hog = vl_hog(img,step_size);
+            %[x,y,z] = size(hog);
+            %hog_features = (reshape(hog,[x*y,z]))';
+            [~, features] = vl_dsift(img, 'Fast', 'Step', step_size);
+            %to reduce precision to 32
+            features = single(features);
+            size(features);
+            [indices, distances] = knnsearch(vocabulary, features');
+
+            %generating histogram and then normalizing it
+            imhist=histc(indices, 1:vocab_size);
+            imhist_norm=imhist./numel(imhist);
+            image_features(image_counter,:) = imhist_norm';      
+            image_counter = image_counter + 1;
+            counter = counter + 1;
+        else
+            counter = counter + 1;
+        end    
+end
+
+fprintf('Computing Bag of words : Negatives\n');
+counter2 = 1;
+
+% for non_vehicle_dataset
+image_counter = 1;
+while(image_counter <= total_images_negative)    
+        filename = strcat(non_plane_source,num2str(random_numbers(counter2)),'.jpg');
+        if exist(filename,'file')
+            if (mod(image_counter,100)==0)
+                fprintf('      ..Processed %d Images\n',image_counter);
+                %fprintf(' Size of hog features : \n');
+                %size(hog_features)
+            end
+            try
+            img = im2single(rgb2gray(imread(filename)));
+%             img = imresize(img,0.2);
+            catch
+                fprintf('Probelm in file %d',counter2);
+                counter2 = counter2 + 1;
+                continue
+            end
+            %fprintf('Processed %d Images\n',image_counter);
+            %hog = vl_hog(img,step_size);
+            %[x,y,z] = size(hog);
+            %hog_features = (reshape(hog,[x*y,z]))';
+            %to reduce precision to 32
+            [~, features] = vl_dsift(img, 'Fast', 'Step', step_size);
+            features = single(features);
+
+            %Find the nearest cluster center in vocabulary for each local feature
+            %in the image based on the Euclidean distance
+            %[indices, distances] = KNNSEARCH(vocabulary,features) returns a vector
+            %   distances containing the distances between each row of features and its
+            %   closest point in vocabulary. Each row in 'indices' contains the index of
+            %   the nearest neighbor in vocabulary for the corresponding row in features.
+            [indices, distances] = knnsearch(vocabulary, features');
+
+            %generating histogram and then normalizing it
+            hist=histc(indices, 1:vocab_size);
+            hist_norm=hist./numel(hist);
+            image_features2(image_counter,:) = hist_norm';
+
+            image_counter = image_counter + 1;
+            counter2 = counter2 + 1;
+        else
+            counter2 = counter2 + 1;
+        end
+end
+        fprintf('Size image features : %d\n', size(image_features,1));
+        fprintf('Size image features2 : %d\n', size(image_features2,1));
+        image_features = [image_features; image_features2];
+        fprintf('Size image features again: %d\n', size(image_features,1));
+end        

detection.m

@@ -0,0 +1,94 @@
+clc
+clear
+close all
+
+run('D:/Softwares/vlfeat-0.9.20/toolbox/vl_setup');
+files = dir('D:\Mandeep\Summer\BTP\Tracking\negatives_training_data\*.jpg');
+count = length(files);
+files2 = dir('D:\Mandeep\Summer\BTP\Tracking\HOG\negatives_training_data\*.jpg');
+count2 = length(files2);
+
+load('weights.mat');
+load('offsets.mat');
+load('vocabulary.mat');
+
+filename = '77.jpg';%1 | 2 | 352| 447 |654 | 77 |
+img = imread(filename);
+original = img;
+%imshow(img);
+
+
+%i - traversing left to right || j - traversing top to bottom || i_n -
+%image number
+i = 1; j = 1;
+% while( (i <= max_y - s_w) && (j <= max_x - s_h))
+i_n = 1;
+bbox ={};
+scores = [];
+images_tried = 0;
+for scale = 1:-0.1:0.6
+    img = imresize(original,scale);
+    %max_y - width of image || max_X - height of image
+    max_y = size(img,2); max_x = size(img,1);
+    %s_h - scale_height || s_w - scale_width
+    s_h = 20;s_w = 45;
+    for i = max_x/2:10:(max_x/2 + max_x/4)
+        for j = 1:10:max_y - s_w
+            %c_i - cropped_image
+            crop = [j,i,s_w,s_h] ;
+            c_i = imcrop(img, crop);
+            %c_i = imread('D:\Mandeep\Summer\BTP\Tracking\Hog\test_data3\1762.jpg');
+            features = extract_features(c_i,vocabulary);
+            [plane,score] = classify(features,weights,offsets);
+            if(plane)
+                i_n
+                scores = [scores;score];
+                crop = [(crop(1)/scale), (crop(2)/scale), (crop(3)/scale),(crop(4)/scale)];
+                bbox{i_n,1} = crop;
+                i_n = i_n + 1;
+                count2 = count2 + 1;
+                filename = strcat('D:\Mandeep\Summer\BTP\Tracking\Hog\negatives_training_data\',num2str(count2),'.jpg');
+                imwrite(c_i,filename);
+            end
+            plane = 0;
+            images_tried = images_tried + 1;
+        end
+    end
+    images_tried
+    images_tried = 0;
+end
+bbox = cell2mat(bbox);
+
+[selectedBbox,selectedScore] = selectStrongestBbox(bbox,scores,'RatioType','Min','OverlapThreshold',0.2);
+img = original;
+for i = 1:size(selectedBbox,1)
+    img = insertShape(img,'rectangle',selectedBbox(i,:),'LineWidth',1);
+
+end
+imshow(img);
+
+function [features] = extract_features(img,vocabulary)
+    img = im2single(rgb2gray(img));
+    vocab_size = size(vocabulary, 1);
+    [~, features] = vl_dsift(img, 'Fast', 'Step', 4);
+    features = single(features);
+    [indices] = knnsearch(vocabulary, features');
+    imhist=histc(indices, 1:vocab_size);
+    imhist_norm=imhist./numel(imhist);
+    features = imhist_norm';    
+end
+
+function [plane,score] = classify(features,weights,offsets)
+    training_score = [];
+    for i = 1:2
+        training_score = [training_score; weights{i}'*features' + offsets{i}];
+    end
+    [~,label_indices] = max(training_score);
+    if (label_indices == 2)
+        score = training_score(label_indices);
+        plane = 1;
+    else
+        plane = 0;
+        score = 0;
+    end
+end

evaluation.m

@@ -0,0 +1,54 @@
+clc
+clear
+close all
+
+run('D:/Softwares/vlfeat-0.9.20/toolbox/vl_setup');
+files = dir('D:\Mandeep\Summer\BTP\Tracking\HOG\negatives_test_data\*.jpg');
+files2 = dir('D:\Mandeep\Summer\BTP\Tracking\HOG\negatives_training_data\*.jpg');
+count = length(files);
+count2 = length(files2)
+
+load('weights.mat');
+load('offsets.mat');
+load('vocabulary.mat');
+i_n = 0;
+for i = 1:6000
+    filename = strcat('D:\Mandeep\Summer\BTP\Tracking\Hog\negatives_test_data\',num2str(i),'.jpg');
+    img = imread(filename);
+    features = extract_features(img,vocabulary);
+    [plane,score] = classify(features,weights,offsets);
+    if(plane)
+        i
+        i_n = i_n + 1;
+        count2 = count2 + 1;
+        filename = strcat('D:\Mandeep\Summer\BTP\Tracking\Hog\negatives_training_data\',num2str(count2),'.jpg');
+        imwrite(img,filename);
+    end    
+end
+
+
+function [features] = extract_features(img,vocabulary)
+    img = im2single(rgb2gray(img));
+    vocab_size = size(vocabulary, 1);
+    [~, features] = vl_dsift(img, 'Fast', 'Step', 4);
+    features = single(features);
+    [indices] = knnsearch(vocabulary, features');
+    imhist=histc(indices, 1:vocab_size);
+    imhist_norm=imhist./numel(imhist);
+    features = imhist_norm';    
+end
+
+function [plane,score] = classify(features,weights,offsets)
+    training_score = [];
+    for i = 1:2
+        training_score = [training_score; weights{i}'*features' + offsets{i}];
+    end
+    [~,label_indices] = max(training_score);
+    if (label_indices == 2)
+        score = training_score(label_indices);
+        plane = 1;
+    else
+        plane = 0;
+        score = 0;
+    end
+end