Multi-pose face recognition by dynamic loss weights

Abstract

This paper explores a face recognition method that combines the dynamic changes of loss weights with face poses. As opposed to current techniques which either expect a single model to learn a variety of poses through massive amounts of training data, or normalize images to a single frontal pose, our method explicitly tackles pose variations by using multiple poses specific models, not only the identity information, but also the facial pose information of the training data is fully utilized. Firstly, we use the solvePnP function in opencv library to classify the facial poses in training dataset. Secondly, we set different loss weights according to the different pose labels. Thirdly, we propose to take into account pose variability by training multiple pose specific models. We leverage deep Convolutional Neural Network (CNN) to learn models with discriminative facial features which we call Dynamic Loss Weights Models(DLW models). We test our models respectively in the LFW dataset and our multi-pose face dataset which contains 50 number of identities. The experiment results show that our designed loss function can improve the performances of profile face recognition.