Learning from summaries of videos: Applying batch mode active learning to face-based biometrics

Against the backdrop of growing concerns about security and privacy, biometrics has emerged as a methodology to reliably infer the identity of an individual. In biometric applications like face recognition, real world data is usually generated in batches such as frames of video in a capture session. The captured data has high redundancy and it is a significant challenge to select the most promising instances from this superfluous set for training a classifier. Active learning methods select only the salient instances for annotation and have gained popularity to reduce the number of examples required to learn a classification model. Typical active learning techniques select one example from an unlabeled set at a time and the classifier is retrained after every selected example. However, there have been very limited efforts in this field to select a batch of salient instances at one shot to update the classification model. In this work, a novel batch mode active learning scheme, specifically tailored to handle the high redundancy of biometric data, has been formulated and validated on the person recognition problem. The instance selection is based on numerical optimization of an objective function, which can be adapted to suit the requirements of a particular application and to integrate additional available information. The results obtained on the widely used VidTIMIT and the NIST MBGC datasets certify the potential of this method in being used for real world biometric recognition problems.