Team Sock Monkeys : CIS 520 Final Report

We experimented with several feature selection and reduction techniques, such as PCA, K-means, and Gabor filtering for images, and mutual information and td-idf encodings for blogs. Additionally, we compared the accuracy of different feature spaces and classification algorithms such as SVMs, KNN, Naive Bayes, and Boosting. Our experiments suggest that selecting good features has the largest effect on accuracy. Lastly, we learned that organic chocolate can be very tasty and comes in many different flavors. 1. Age and Gender prediction from facial images We looked at two different aspects of the problem for this task: (a) the feature space to use and, (b) the classification algoriths to use. Most background literature [4] used a tightly fitting elliptical region around the face (excluding any hair and background context) to train the classifiers. For our training data, this kind of cropping was not available though it could be enforced by applying an off-the-shelf face localizer to the data. However, rather than defining such a crop, we experimented with the opposite approach of including more contextual information around the face. Thus, for all the experiments below, we first cropped the images to an area bigger than the original crop in the baseline code and it resulted in better performance from our crossvalidation experiments. In addition, the cropped area was scaled down by a factor of 4 unlike the fixed 25×25 square in the baseline code. All the algorithms described below were tried for both age and gender prediction initially so some of them use both gender and age labels together during training. Later, for each task, we chose the specific variant which performed better for that task without removing the dependence on both age and gender labels. We think that the joint labels might have helped in certain cases depending on the correlation between age-related changes and gender. Figure 1. Cluster centers learned for the male and female genders (rows 1 and 2), and the three age categories (rows 3-5).