This tutorial will address the emerging research area of heterogeneous face recognition (HFR), which seeks to match facial probe imagery acquired in one modality against a gallery database of facial images acquired in another modality. HFR has strong potential to provide new capabilities for law enforcement, the military, and the intelligence community. In this tutorial, we will focus on infrared-to-visible face recognition for low-light and nighttime applications, discussing feature extraction techniques, regression methods, and classification algorithms for matching infrared imagery to gallery databases containing only visible imagery. We will also present recent advances in exploiting sensor technology such as polarimetric imaging for HFR, and discuss new algorithms such as generative adversarial network based approaches for HFR.
Presenters: Sean Hu, Ben Riggan, Nathan Short, and Vishal Patel
In the last years, many biometric recognition algorithms have been proposed, and many biometric datasets are used by researchers to compare their results. However, even when datasets provide an evaluation protocol that assures comparable results, researchers need to implement this protocol on themselves. Hence, instead of focusing on the algorithm to be tested, much time is spent to implement the framework.
In this tutorial we will introduce the biometric recognition framework of Bob. This framework includes both the implementation of evaluation protocols for several publicly available datasets, as well as implementations of many biometric recognition algorithms, while it is still easy to extend. This framework makes it easy to provide reproducible research, and several reproducible research papers have been published using Bob and its biometric recognition framework.
On the basis of face recognition examples, participants will be made familiar with the structure of a biometric recognition experiment including concepts of feature extraction, model enrollment, score computation, and evaluation. In three hands-on exercises, the participants will run three very different face recognition algorithms. After starting with a simple Eigenface approach, the more complex algorithm of Gaussian mixture modeling of DCT features will be introduced and evaluated. Finally, we will extend the framework by introducing a new algorithm, i.e., extracting deep features from the publicly available VGG face network using the Python interface of Caffe.
Presenter: Manuel Gunther