Generalization on unseen domains via inference-time label-preserving target projections

Published in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021

We propose a novel method for domain generalization using an inference-time optimization procedure. We first learn a label-sensitive, domain-agnostic metric over images. We then learn a generative model over the resulting representations. Finally, during inference, we optimize over the latent space of the generative model to generate a point from the source domain which preserves the target label.