Min Je Jung

Project 5 / Face Detection with a Sliding Window

In this project, I implemented the sliding window detector as written in a paper by Dalal and Triggs. There were four functions we had to implement for this project:

1. Getting Positive Features
2. Getting Random Negative Features
3. Classifier Training
4. Running the Detector

Getting Positive Features

I get each image and compute the Histogram of Oriented Gradients using vl_hog.

Getting Random Negative Features

For a random image, I grayscaled it. Then I get random samples of the image and use vl_hog to find the HOG and add it on to the negative training examples.

Classifier Training

I call vl_svmtrain for the classifier training from the data vectors, positive and negative features, and the labels, 1's and -1's, and the lambda as .0001.

Running the Detector

I basically convert each image to HOG feature space using vl_hog for each scale. I take the Hog spaces and create a space with them. I then use a confidence level of .75 to find greater than that.

Example Code: classifier_training

function [w, b] = classifier_training(features_pos, features_neg)

	features_pos = features_pos';
	[~, numbers_pos] = size(features_pos);

	features_neg = features_neg';
	[~, numbers_neg] = size(features_neg);

	labels = ones(1, (numbers_pos + numbers_neg));
	labels(numbers_pos + 1 : end) = -1 * labels(numbers_pos + 1 : end);

	[w, b] = vl_svmtrain([features_pos, features_neg], labels, 0.0001);

end

Results in a table

Face template HoG visualization learned after training data. It looks roughly like a face after classifier training.

Precision Recall curve for the test data.