Template-based Unconstrained Face Recognition

As an important biometric identification technology,face recognition is widely used in many practical applications.Thanks to deep learning,the state-of-the-art algorithms have approached more than 99%accuracy in LFW dataset.However,these achievements are only suitable for constrained face recognition scenarios where faces have near frontal bias.Unconstrained face recognition is still quite difficult for two reasons.On the one hand,faces in this setting are more unconstrained in head pose,yaw angle,illumination and expression,causing large variance within each subject,even larger than the inter-subject difference.On the other hand,unconstrained face verification is performed on set-to-set level rather than image-to-image level.A set of images and videos of a certain person is called a template.In this paper,we mainly focus on template-based unconstrained face recognition.In order to tackle the problems as described above,Attention-Based Template Adaptation(ABTA)is proposed in this paper,which suits the application setting of template-based unconstrained face verification.The algorithm takes VGG-FACE as the basic neural network architecture,and utilizes advanced pose-aware alignment,Triplet Probabilistic Embedding(TPE),attention-based Neural Aggregation Network(NAN)and Template Adaptation(TA)algorithms successively.Firstly,unconstrained faces are aligned based on their pose,which reduces the pose variance of the faces to a certain extent.Then,VGG-FACE is used to extract features for all the faces in the templates.After that,TPE is applied to do metric learning and dimensionality reduction at the same time,with metric learning reducing the data bias between training data and testing data and dimension reduction resulting in more compact feature representation.In order to simplify the complexity of template-based verification,we apply Neural Aggregation Network(NAN)to fuse the face features in each template,so that each template corresponds to a compact fixed length feature representation.The Neural Aggregation Network is attention-based,which promotes the beneficial faces while discarding noisy information simultaneously.Finally,Template Adaptation(TA)is applied to transfer knowledge learned from the background set to testing set via one-vs-all support vector machine(SVM),which significantly improves the performance of the algorithm.The ABTA algorithm proposed in this paper is an improvement on the Template Adaptation integrating some other sophisticated methods.The contribution of this paper is three-fold:1)applying pose-aware alignment to reduce pose variance within similar pose,which enables the algorithm to focus on the difference between faces rather than the pose;2)using the features produced by TPE as the input features to Template Adaptation,which is more discriminative and compact(4096-d vs 512-d)than original VGG features;3)utilizing attention-based features generated by NAN rather than the mean features of each templates as the template centers used in Template Adaptation.The original Template Adaptation is slow and the features is less discriminative.The ABTA algorithm proposed in this paper has some advantages due to its compact and discriminative features.The ABTA algorithm produces results comparable to the state-of-the-art in the challenging face dataset,IJBA,with faster inference time.