A hierarchical training and identification method using Gaussian process models for face recognition in videos

In a video based face identification task, a sequence of frames can be utilized to identify the subject in the video. The information extracted from frames can provide samples of the subject in different head poses and facial expressions and under various lighting conditions which enriches the training process. However, some of these frames may not be useful for identification due to noise from various sources (such as occlusion, low resolution, and face tracking errors). It is important to reduce the effect of noisy samples by designing a representation structure that is capable of alleviating the noise in each sequence, complemented by developing a recognition procedure that rejects the wrong decisions affected by noise. In this paper we propose a sequence representation called Ensemble of Abstract Sequence Representatives (EASR) that is aimed at reducing the effect of noisy frames in a sequence. EASRs are used to guide the sampling process in a learning scheme called specialization - generalization which is used to train an ensemble of binary Gaussian Process (GP) models. Identification is done using: (i) the similarity between the EASRs of the gallery and probe images, and (ii) the label provided by the ensemble of GP classifier models. Evaluation of our approach on three publicly available benchmark datasets demonstrates significantly better performance compared to the state-of-the-art.