Invited Talk

Title of the Talk: How to Detect When the Probe in 1-to-N Facial ID is Out-of-Gallery
Speaker: Kevin W. Bowyer, University of Notre Dame, Indiana
Abstract: 1-to-N face matching is the algorithmic engine in finding an investigative lead using a probe image from a frame of surveillance video. When the person in the probe image is not in the gallery, any result from 1-to-N facial ID has to be a false identification. Detecting out-of-gallery probes has potential to reduce wasted investigative effort and reduce wrongful arrests. We discuss new approaches to detecting when a 1-to-N facial ID is out-of-gallery. Our approaches do not use the similarity score between images. We evaluate the accuracy of our approaches on probe images that are blurred or low resolution, as well as on mugshot-quality probes.
Bio: Kevin Bowyer is the Schubmehl-Prein Family Professor Emeritus at the University of Notre Dame. He is a Fellow of the AAAS “for distinguished contributions to the field of computer vision and pattern recognition, biometrics, object recognition and data science”, a Fellow of the IEEE “for contributions to algorithms for recognizing objects in images”, and a Fellow of the IAPR “for contributions to computer vision, pattern recognition and biometrics”. Professor Bowyer has served as Editor-In-Chief of both the IEEE Transactions on Pattern Analysis and Machine Intelligence and the IEEE Transactions on Biometrics, Behavior, and Identity Science. Professor Bowyer has also served as General or Program Chair of conferences such as Computer Vision and Pattern Recognition, Winter Conference on Applications of Computer Vision, Face and Gesture Recognition, and the International Joint Conference on Biometrics.

Title of the Talk: Face, Body, and Person Identification in Real-world Viewing Conditions
Speaker: Alice O’Toole, University of Texas, Dallas
Abstract: Face recognition algorithms are highly accurate at establishing the unique identity of individuals in controlled conditions. In natural viewing conditions, however, facial identity information is commonly degraded or obscured (e.g., viewed at distance or from extreme angles). When the face is unusable or inaccessible, information about the shape of the body can constrain identity decisions. Body shape can contribute to person identification by supporting or vetoing tentative face identifications. As such, it can serve as a valuable biometric, even if it does not uniquely identify an individual. In this talk, I will begin with what we have learned about representations of faces in deep networks. I will then present multiple machine-based approaches to body/person identification. I will also explore the complex challenge of integrating face and body information to achieve more accurate person identification. I will draw on lessons learned from the human visual system, which accomplishes this integration with remarkable flexibility and adaptability, modulating its reliance on the face vs. body depending on the viewing conditions.