Dynamic ensembles of exemplar-SVMs for still-to-video face recognition

摘要: An efficient multi-classifier system is proposed for robust still-to-video FR.Multiple diverse representations are generated from the single target still face.Individual-specific ensembles of exemplar-SVM designed based on domain adaptation.Different adaptation training schemes to generate classifiers.Dynamic classifier selection and weighting applied perform spatio-temporal FR. Face recognition (FR) plays an important role in video surveillance by allowing accurately recognize individuals interest over a distributed network cameras. Systems FR exposed challenging operational environments. The appearance faces changes when captured under unconstrained conditions due variations pose, scale, illumination, occlusion, blur, etc. Moreover, facial models used matching may not be intra-class because they typically priori with one reference per person. Indeed, during enrollment (using cameras) differ considerably those operations cameras). In this paper, (MCS) accurate multiple face (DA). individual-specific ensemble (e-SVM) classifiers thereby improve robustness variations. During individual, model still, where descriptors random feature subspaces allow pool patch-wise classifiers. To adapt these domains, e-SVMs trained using labeled patches extracted versus cohort other non-target stills mixed unlabeled corresponding trajectories operations, most competent given probe dynamically selected weighted internal criteria determined space e-SVMs. This paper also investigates impact different DA, as well as, validation set unknown domain. performance was validated videos COX-S2V Chokepoint datasets. Results indicate that can surpass state-of-the-art accuracy, yet significantly lower computational complexity. dynamic combine only relevant each input probe.