Colloquium 1 - Oct. 25 : Ross Beveridge

Automated Face Recognition at CSU and at Large
Ross Beveridge

This colloquium, presented by Ross Beveridge of Coloado State University, discussed the current approaches to face recognition, as well as the pitfalls, successes and open issues surrounding research in the area, particularly as they relate to the work being done with the CSU Face Indentification System. Core algorithms were presented at the talk which included the PCA/Eigenfaces algorithm, the hybrid PCA+LDA algorithm, the Bayesian Intrapersonal/Extrapersonal Classifier and the Elastic Bunch Graph Matching.

Ross presented a few of the issues concerning the algorithms and how they dealt with the false turn away problem - that is rejecting a correct face. He noted in particular that the sample imagery used was very important, for example, changing the location from controlled background and lighting to a hallway increased false turn away from 1 in 100 faces to 20 in 100 faces, demonstrating that the current algorithms are very easy to break.

Since face recognition is most often used in environments where security, access and authorization are important, he noted that when 2 sources of data were used, also called the Data Fusion Method, success was improved even more. For example, Data Fusion takes an image (of a face) and text together and computes a match score. The text is analyzed using a text classifier algorithm. These experiments were done with Wikipedia and Yahoo! text and images. Ross noted this was an interesting experiment to follow, but was difficult to replicate in a real-time environment, and of course was particularly difficult when sample images used varied greatly.

Ross continued to talk about the nature of various algorithms and how facial features were used as anchors within the algorithms. One stunning statement he made, however, was that "raw" pixel analysis beats or ties the 2nd best feature and algorithm combination! He went on to talk about how the quality of images affects the end results, and the high resolution images are a benefit to overall accuracy. Furthermore, he suggested neutral smiling actually decreased the accuracy of results - surprisingly smiling was a benefit to the correctness of the results.

This talk was very interesting, though I was surprised to find a large number of constraints are still required to get good results. I was also surprised that available computing power had not provided high enough gains to provide better results to the algorithms being used.