Ensembles of exemplar-SVMs for video face recognition from a single sample per person

摘要: Recognizing the face of target individuals in a watch-list is among most challenging applications video surveillance, especially when enrollment based on one reference still facial image. Besides limited representativeness models used for matching, appearance faces captured videos varies due to changes illumination, pose, scales, etc., and camera inter-operability. A multi-classifier system proposed this paper robust still-to-video recognition (FR) multiple diverse representations. An individual-specific ensemble exemplar-SVMs (e-SVMs) classifiers assigned each person, where classifier trained using high-quality versus many lower-quality non-target videos. Diverse representations are generated from different patches isolated images descriptors that various nuisance factors (e.g., illumination pose) commonly encountered surveillance environments. Discriminant feature subsets, training samples, fusion functions selected scene. Experiments Chokepoint dataset reveal e-SVMs outperforms state-of-the-art FR systems specialized single sample per person problem.