Step 3 - File edited: build_vocabulary.m

This file is used to build the vocabulary database to construct the histogram during the later task. I collected 100 sift features from each image and tried step size of both 10 and 20. I finally decided to keep step size as 20 due to runtime concern. Since the step size in get_bags_of_sift.m is supposed to be smaller or equal to the step size of build_vocabulary.m, using 20 for build_vocabulary allows me to amplify the step size in get_bags_of_sift.m to reduce computations.

Step 4 - File edited: get_bags_of_sifts.m

This file helps us to construct the histogram that explains the features of the test images. Likewise, sift features are collected from each test images and compared against the vocabulary we built ealier. For each feature from test image, we find the closest feature match it has on the vacabulary and adds 1 to each bin. We normalize the histograms and return it. I chose step size of 10 in this file as it seems to give me the highest accuracy rate.

When bags of sifts and nearest neighbor classifer are used, I achieved accuracy of 52.9%.

Step 5 - File edited: svm_classify.m

categories = unique(train_labels);
num_categories = length(categories);
predicted_categories = cell(size(test_image_feats,1),1);
confidence_matrix = zeros(size(test_image_feats,1),num_categories);
for i=1:num_categories
    indices = strcmp(categories(i),train_labels);
    train_binary = zeros(length(indices),1);
    for l=1:length(indices)
        if indices(l) == 0
            train_binary(l) = -1;
        else
            train_binary(l) = 1;
        end
    end
    [W,B] = vl_svmtrain(train_image_feats', train_binary, 0.0003);
    for j=1:size(test_image_feats,1)
        confidence = dot(W,(test_image_feats(j,:))') + B;
        confidence_matrix(j,i) = confidence;
    end
end

Above is part of the code from the file. For each unique category, i selected pictures of that category from training samples and set their label to 1. Then I constructed SVM for each category and calculate the score of each test picture using that SVM. Eventually, each test image will get number of scores equal to the number of categories and select the category that seems to most likely fit the test image. For the choice of lambda, I tried multiple numbers, and finally set the lambda = 0.0003 as it renders the highest accuracy rate.

Results in a table

lambda	Feature selection	Classifier	Accuracy	Runtime
	Tiny images	Nearest Neighbor	19.1	pretty short
	Bags of SIFT (vocab step = 10,bags of sift step = 5)	Nearest Neighbor	53.3	17 mins
	Bags of SIFT (vocab step = 10,bags of sift step = 10)	Nearest Neighbor	50.8	12 mins
	Bags of SIFT (vocab step = 20,bags of sift step = 10)	Nearest Neighbor	52.9	8 mins
0.00001	Bags of SIFT (vocab step = 10,bags of sift step = 5)	SVM	66.1	12 mins
0.0001	Bags of SIFT (vocab step = 10,bags of sift step = 5)	SVM	68.3	12 mins
0.001	Bags of SIFT (vocab step = 10,bags of sift step = 5)	SVM	60.5	12 mins
0.0005	Bags of SIFT (vocab step = 10,bags of sift step = 5)	SVM	65.1	12 mins
0.0001	Bags of SIFT (vocab step = 20,bags of sift step = 10)	SVM	62.7	8 mins
0.0005	Bags of SIFT (vocab step = 20,bags of sift step = 10)	SVM	61.7	8 mins
0.0003	Bags of SIFT (vocab step = 20,bags of sift step = 10)	SVM	63.3	8 mins