Joel's Colloquium 2

Ross Beveridge

Beveridge spoke about face recognition algorithms being developed at CSU and abroad. He spoke in length about the CSU Face Identification Evaluation System which compares 4 different face recognition algorithms: Principle Components Analysis, Principle Components Analysis and Linear Discriminate Analysis, Bayesian Interpersonal/Extrapersonal Classifier, and Elastic Bunch Group Matching. All of these algorithms are written in C, and the preprocessing code replicates that in the FERET evaluations. The hope is that people will download their system and compare novel face recognition algorithms to these four baseline algorithms through a standard interface and standard tests.

Beveridge went into a little detail about the Elastic Bunch Group Matching algorithm. The way it works is to find landmark features in an image, then finding similarities between these landmarks. Beveridge indicated that this algorithm is currently the most accurate facial recognition algorithm.

Beveridge spoke briefly about the different types of recognition, from the "Find the best" question, where the computer searches a database to find the closest match, to "Am I who I say I am", where it verifies that a person is who they claim to be. Currently, slight changes such as smiling can completely fool a recognition algorithm. An interesting point is that government agencies require that employees not smile for their ID picture, and Beveridge points out that smiling actually makes computerized facial recognition easier.