Competitions in Education: Case Study on Face Verification

Abstract

All genuine knowledge originates in direct experience, especially for engineering courses. To help the students grasp hands-on experience of solving practical problems, a Machine Learning competition named TUGraz-TUT Face Verification Challenge was jointly organized by Graz University of Technology and Tampere University of Technology. The objective of the competition was to identify whether two facial images represent the same person. During the two-month period, the competition received 137 entries submitted by 28 players in 20 teams. This thesis summarizes the outcome of the competition. To scrutinize the face verification system systematically, the processing workflow was divided into several parts. In the procedure of face alignment, Unsupervised Joint Alignment and Ensemble of Regression Trees were compared. Subsequently, the OpenFace and VGG Face features were retrieved from the aligned images. In the classification system, the performance of neural network and support vector classification were evaluated. Moreover, the influence of the ensemble strategies and the result of different error metrics were investigated. Based on the cutting-edge deep neural networks proposed by the research community, the winning solutions attained excellent results as the Weighted AUC scores exceeded 0.9990. In addition to the preceding accomplishments, the findings suggested that there were still opportunities for further enhancements of the face verification systems. The limitations of current work and a handful of conceivable directions for future research had been deduced.