Meta-Recognition: Score Analysis and Calibration for Recognition Problems: Walter J. Scheirer, Harvard
- 1:00 pm on Monday, April 8, 2013
- 2:00 pm on Monday, April 8, 2013
- MCS 148
Abstract: Recent work has shown that visual attributes are a powerful approach for applications such as recognition, image description and retrieval. However, fusing multiple attribute scores – as required during multi-attribute queries or similarity searches – presents a significant challenge. Scores from different attribute classifiers cannot be combined in a simple way; the same score for different attributes can mean different things. In this talk, I will show how to construct normalized “multi-attribute spaces” from raw classifier outputs, using a technique called “Meta-Recognition,” which is based on the statistical Extreme Value Theory. This method calibrates each raw score to a probability that the given attribute is present in the image. These probabilities can be fused in a simple way to perform more accurate multi-attribute searches, as well as enable attribute-based similarity searches. A significant advantage of this approach is that the normalization is done after-the-fact, requiring neither modification to the attribute classification system nor ground truth attribute annotations. We’ll look at some image retrieval results for a large data set of nearly 2 million face images, as well as a study of perceptual similarity of attribute combinations based on distances in a normalized classifier space. Meta- Recognition is a general theory, applying to any recognition system producing distance or similarity scores. With that in mind, we’ll also look at examples for object recognition, biometric recognition, and support vector machines as the overall attribute model is developed. Bio: Dr. Walter J. Scheirer received his Ph.D. (engineering, with a concentration in computer science) from the University of Colorado, Colorado Springs, and his M.S. (computer science) and B.A. (computer science and international relations) degrees from Lehigh University. He is currently a Postdoctoral Research Fellow in the departments of Computer Science and Molecular and Cellular Biology, and Center for Brain Science at Harvard University. He also holds the position of Assistant Professor Adjoint in the Department of Computer Science at the University of Colorado, Colorado Springs. Prior to coming to Harvard, Dr. Scheirer was the Director of Research and Development at Securics, Inc., an early-stage technology company specializing in innovative biometric security solutions. Dr. Scheirer is a noted expert in the area of human biometrics, and has published and lectured widely on a variety of topics, including computer vision, machine learning, pattern recognition, and digital humanities. His current research work focuses on multi-classifier fusion, machine learning for open set recognition, and unconstrained face and object recognition. At Securics, Dr. Scheirer helped develop the first privacy preserving biometric key infrastructure for network transactions with an identity component.