CrossEncoder: Towards 3D-free depth face recovery and fusion scheme for heterogeneous face recognition

Abstract

As a worthwhile trade-off between fully 2D and fully 3D based face recognition (FR), 2D/3D asymmetric face recognition has recently emerged which involves matching two face representations from these two alternative modalities. Different from most previous work which rely on accurately registered 3D face templates, we address this issue by proposing a novel multi-task deep convolutional neural network (CNN) architecture based on 2.5D images. With an autoencoder-like pipeline and a specially formulated criterion, our approach is capable of parallelizing real-time 2.5D face image reconstruction and discriminative face feature extraction. Further, through both qualitative and quantitative experiments on commonly used FRGC 2D/3D face database, we demonstrate that our framework could achieve satisfactory performance on both tasks while being drastically efficient.