CoConGAN: Cooperative contrastive learning for few-shot cross-domain heterogeneous face translation

The inherently uneven informative setting in heterogeneous face images makes heterogeneous face translation challenging for synthesizing the analogous face image which preserves the identity of the input image from the source domain and fits the style of the reference image from the target domain. While effective, current methods require access to images with the same identity in both the source and target domains at training time. However, this is a rigid restriction, and the identities from the source domain may not be seen in the target domain, which limits the performance of translation and heterogeneous face recognition. Motivated by the human capability of picking up the essence of a novel face from a small number of examples and generalizing from there, we seek a few-shot, unsupervised diverse face-to-face translation framework that works on previously unseen identities that are specified, at training and testing time, only by a few example images. Besides, two issues have to be solved in this framework: the identity consistency and face diversity. We therefore propose a novel cooperative contrastive loss, which can jointly encourage identity diversity and preserve identity consistency. Through extensive experimental validation and comparisons to several baseline methods on benchmark datasets, we verify the effectiveness of the proposed framework.