Casey Barnette

image_feats = zeros(size(image_paths,1), 256);
for i = 1:size(image_paths,1)
    image_feats(i,:) = reshape(imresize(imread(image_paths{i}), [16 16]), [1 256]);
    %normalize
    image_feats(i) = image_feats(i)/norm(image_feats(i));
end

Bag of Sift

This is the other feature extracting part of our assignment. We start by building a vocabulary of size 200. We do this by grabbing sift features for every image with a relatively large step size, for my project I chose a step size of 20. Then we take all of the sift features and perform kmeans clustering on the features with the number of clusters equal to our vocab size. Then once that is built we can actually build our bags of sift features. For this we grab the sift features for each image, for my best results I used a step size of 5 with fast sift off and for the faster results I used a step size of 10 with fast sift on. Then we bin the sift features into their appropriate cluster which creates a histogram of counts for each cluster. Then I normalize the histogram in order to keep images with a large amount of sift features from over powering the smaller ones.

Nearest Neighbor

This is the more naive of the two classifiers that we created. For this part I implemented single nearest neighbor. To do this I calculated the minimum distance for each for each test feature to each training feature. I then take the label from the training feature that is closest to the test feature and assign it to that test feature. I attempted to do k-nearest neighbor as can be seen in the commented out code; however, it seemingly made no difference to the overall accuracy.

Support Vector Machines

This task was the smarter classifier of the two that we created. For this I created a vector of 1 vs All SVMs using vl_svmtrain.

models = cell(num_categories, 1);
biases = cell(num_categories, 1);
for i = 1:num_categories
    labels = -1*ones(size(train_image_feats,1), 1);
    matching = strcmp(categories{i}, train_labels);
    labels(matching) = 1;
    [model, bias] = vl_svmtrain(train_image_feats', labels', lambda);
    models{i} = model;
    biases{i} = bias;
end

I tried this with many different lambdas from 0.00000001 to 1 and found that 0.00001 provided the highest accuracy in all cases for me. This is likely due to underfitting/overfitting problems. The larger lambda is the closer you get to having an underfitting problem and the smaller lambda gets the closer you are to having an overfitting problem. This is because lambda controls how many different iterations the SVM undergoes when trying to divide the hyperplane.

Results

Best results with time disregarded. SVM + BOS step size 5

Accuracy is 0.667

Results for SVM and Bags of Sift with time constraints. Step size 10 and fast flag.

Accuracy is 0.562

Results for KNN and Bags of Sift with time constraints. Step size 10 and fast flag.

Accuracy is 0.521

Results for KNN and Tiny Images

Accuracy is 0.193

Category name	Accuracy	Sample training images		Sample true positives		False positives with true label		False negatives with wrong predicted label
Kitchen	0.080					Industrial	Forest	Highway	Forest
Store	0.040					Office	TallBuilding	TallBuilding	Coast
Bedroom	0.200					InsideCity	Store	OpenCountry	Mountain
LivingRoom	0.100					Office	Kitchen	Bedroom	Office
Office	0.190					Mountain	Bedroom	Industrial	InsideCity
Industrial	0.130					LivingRoom	TallBuilding	InsideCity	Store
Suburb	0.470					Coast	InsideCity	Kitchen	Street
InsideCity	0.030					Suburb	Coast	TallBuilding	Store
TallBuilding	0.180					Store	Forest	LivingRoom	Coast
Street	0.600					Kitchen	Store	Highway	Suburb
Highway	0.750					Suburb	Industrial	Coast	Mountain
OpenCountry	0.310					Forest	Kitchen	Suburb	Coast
Coast	0.400					Mountain	Mountain	InsideCity	OpenCountry
Mountain	0.290					Store	Bedroom	Kitchen	Forest
Forest	0.140					Mountain	Street	Suburb	Kitchen
Category name	Accuracy	Sample training images		Sample true positives		False positives with true label		False negatives with wrong predicted label

Lorem ipsum dolor sit amet, consectetur adipisicing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat. Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur. Excepteur sint occaecat cupidatat non proident, sunt in culpa qui officia deserunt mollit anim id est laborum.