Added MATLAB scripts

getEntropy.m

@@ -0,0 +1,7 @@
+function entropy = getEntropy(p_dist)
+entropy = 0;
+for i = 1:size(p_dist,2)
+    if(p_dist(i) > 0)
+        entropy = entropy - p_dist(i) * log2(p_dist(i));
+    end
+end

readDictionary.m

@@ -0,0 +1,75 @@
+% read YUV Dictionary
+% clear; close all; clc;
+
+% load Dictionary
+% YUV_Dict = load('GoodDictionaryCyberZooPinkMat.dat');
+YUV_Dict = load('Dictionary_00000.dat');
+% YUV_Dict = load('Dictionary_big_flowers.dat');
+
+% params
+n_textons = 20; % number of visual words
+patch_size = 6; % size of one patch
+
+% extract YUV components
+U_ALL = YUV_Dict(1:4:end,1);
+Y1_ALL = YUV_Dict(2:4:end,1);
+V_ALL = YUV_Dict(3:4:end,1);
+Y2_ALL = YUV_Dict(4:4:end,1);
+
+Dictionary = zeros(n_textons,patch_size*patch_size*3);
+
+
+hh = figure();
+
+for w = 1:n_textons
+    subplot(ceil(sqrt(n_textons)), ceil(sqrt(n_textons)), w);
+
+    % extract YUV per texton
+    U = U_ALL((w-1)*18+1:w*18, 1);
+    Y1 = Y1_ALL((w-1)*18+1:w*18, 1);
+    V = V_ALL((w-1)*18+1:w*18, 1);
+    Y2 = Y2_ALL((w-1)*18+1:w*18, 1);
+
+    % conversion
+    R1 = Y1 + 1.4022 .* (V - 128);
+    G1 = Y1 - 0.3456 .* (U - 128) - (0.7145 .* (V - 128));
+    B1 = Y1 + 1.7710 .* (U - 128);
+    R2 = Y2 + 1.4022 .* (V - 128);
+    G2 = Y2 - 0.3456 .* (U - 128) - (0.7145 .* (V - 128));
+    B2 = Y2 + 1.7710 .* (U - 128);
+
+    R = zeros(patch_size,patch_size);
+    G = zeros(patch_size,patch_size);
+    B = zeros(patch_size,patch_size);
+    R(:,1:2:end) = reshape(R1,3,6)';
+    R(:,2:2:end) = reshape(R2,3,6)';
+    G(:,1:2:end) = reshape(G1,3,6)';
+    G(:,2:2:end) = reshape(G2,3,6)';
+    B(:,1:2:end) = reshape(B1,3,6)';
+    B(:,2:2:end) = reshape(B2,3,6)';
+
+    % clip the values into range [0, 255]
+    R = max(0, min(R, 255));
+    G = max(0, min(G, 255));
+    B = max(0, min(B, 255));
+    % R = min(max(R, 0),255);
+    % G = min(max(G, 0),255);
+    % B = min(max(B, 0),255);
+
+    % form bgr image
+    R_norm = R./255;
+    G_norm = G./255;
+    B_norm = B./255;
+    rgbimg (:,:,1) = B_norm;
+    rgbimg (:,:,2) = G_norm;
+    rgbimg (:,:,3) = R_norm;
+
+    Dictionary(w,1:patch_size*patch_size) = reshape(rgbimg(:,:,1)',1,36);
+    Dictionary(w,patch_size*patch_size+1:patch_size*patch_size*2) = reshape(rgbimg(:,:,2)',1,36);
+    Dictionary(w,patch_size*patch_size*2+1:patch_size*patch_size*3) = reshape(rgbimg(:,:,3)',1,36);
+
+    imshow(rgbimg);
+    title(num2str(w))
+end
+
+% save dictionary_visualwords2_rgb.mat Dictionary

train.m

@@ -0,0 +1,154 @@
+function train()
+
+MAX_SAMPLES = 25000;
+start_sample = 1;
+
+BIAS = true;
+PRIOR = true;
+TEST_SET = true;
+test_ratio = 0.30;
+
+weights = false; % whether we have onboard weights we want to compare to the current learning:
+if(weights)
+    w = load('Weights_00000.dat');
+end
+
+% structure A:
+% 1) height
+% 2) gain
+% 3) cov div
+% 4-end) textons
+A = load('Training_set_00000.dat');
+n_samples = size(A,1);
+if(TEST_SET)
+    n_training = round((1-test_ratio) * n_samples);
+    n_test = n_samples - n_training;
+    start_ind = floor(rand(1) * (n_samples - n_test));
+    A_test = A(start_ind:start_ind+n_test, :);
+    A = [A(1:start_ind-1, :); A(start_ind+n_test+1:end, :)];
+    % A_test = A(n_training+1:end, :);
+    % A = A(1:n_training, :);
+end
+
+% **************************************
+% learn a mapping from features to gain:
+% **************************************
+
+A = A(start_sample:min([MAX_SAMPLES, size(A,1)]), :);
+b = A(:,2);
+f = A(:,4:end);
+if(BIAS)
+    AA = [f, ones(size(A,1),1)];
+else
+    AA = f;
+end
+if(~PRIOR)
+    x = AA \ b;
+else
+   alpha = 1; %10;
+   x = inv(AA' * AA + alpha * eye(size(AA, 2))) * AA' * b; 
+end
+
+% store the resulting weights:
+fid = fopen('Weights_MATLAB.dat', 'w');
+for i = 1:length(x)-1
+    fprintf(fid, '%f ', x(i));
+end
+fprintf(fid, '%f', x(end));
+fclose(fid);
+
+% evaluate the resulting estimates and compare them with the weights
+% learned onboard:
+y = AA * x;
+height_gain_estimate = y;
+fprintf('MAE on training set = %f\n', mean(abs(y-b)));
+if(weights)
+    Z = AA * w';
+end
+figure(); plot(y); hold on; plot(b);
+if(weights)
+    plot(Z);
+end
+title('Height on training set')
+legend({'height gain estimate', 'height gain', 'onboard gain estimate'});
+
+figure(); plot(smooth(y, 20)); hold on; plot(b);
+if(weights)
+    plot(smooth(Z, 20));
+end
+title('Smoothed Height')
+legend({'height gain estimate', 'height gain', 'onboard gain estimate'});
+
+figure();
+bar(x, 'FaceColor', [1 0 0]); % hold on; bar(w);
+title('Weights learned in MATLAB');
+
+figure();
+plot(A(:,1)); hold on; plot(A(:,2));
+legend({'Height', 'Gain'});
+
+% MAE on test set:
+f_test = A_test(:,4:end);
+if(BIAS)
+    AA_test = [f_test, ones(size(A_test,1),1)];
+else
+    AA_test = f_test;
+end
+y_test = AA_test * x;
+b_test = A_test(:, 2);
+fprintf('MAE on test set = %f\n', mean(abs(y_test-b_test)));
+if(weights)
+    Z_test = AA_test * w';
+end
+figure(); plot(y_test); hold on; plot(b_test);
+if(weights)
+    plot(Z_test);
+end
+title('Height on test set')
+legend({'height gain estimate', 'height gain', 'onboard gain estimate'});
+
+
+% ***********************************
+% now learn a function to learn sonar 
+% ***********************************
+
+b = A(:,1);
+f = A(:,4:end);
+if(BIAS)
+    AA = [f, ones(size(A,1),1)];
+else
+    AA = f;
+end
+x = AA \ b;
+y = AA * x;
+figure(); plot(smooth(y, 20)); hold on; plot(b);
+
+b = A(:,1);
+f = height_gain_estimate;
+AA = [f, ones(size(A,1),1)];
+x = AA \ b;
+y = AA * x;
+plot(smooth(y,20));
+title('Sonar height')
+legend({'estimate trained with sonar', 'sonar', 'scaled gain estimate'});
+
+figure();
+plot(A(:, 3));
+title('Cov div');
+
+entr = getEntropies(f);
+p_peak = zeros(1, size(f,2));
+p_peak(1) = 1;
+min_entr = getEntropy(p_peak);
+p_uniform = ones(1, size(f,2)) ./ size(f,2);
+max_entr = getEntropy(p_uniform);
+figure();
+histogram(entr, 30);
+title(['Entropies: min = ' num2str(min_entr) ', max = ' num2str(max_entr)]);
+
+function entropies = getEntropies(f)
+n_el = size(f,1);
+entropies = zeros(n_el,1);
+for el = 1:n_el
+   entropies(el) = getEntropy(f(el, :));
+end