Learning One-Shot Exemplar Svm From The Web For Face Verification

We investigate the problem of learning from a single instance consisting of a pair of images, often encountered in unconstrained face verification where the pair of images to be verified contain large variations and are captured from never seen subjects. Instead of constructing a separate discriminative model for each image in the couple and performing cross-checking, we learn a single Exemplar-SVM model for the pair by augmenting it with a negative couple set, and then predict whether the pair are from the same subject or not by asking an oracle whether this Exemplar-SVM is for a client or imposter in nature. The oracle by itself is learnt from the behaviors of a large number of Exemplar-SVMs based on the labeled background set. For face representation we use a number of unlabeled face sets collected from the Web to train a series of decision stumps that jointly map a given face to a discriminative and distributional representation. Experiments on the challenging Labeled Faces in the Wild (LFW) verify the effectiveness and feasibility of the proposed method.