Supervised Domain Adaptation with Disjoint Label Spaces for Fine-Grained Classification

Domain adaptation scenarios commonly assume that the label spaces of the source and target domains are either equal or share a common set of classes. However, in fine-grained classification settings, it is likely that the common label set is empty. Therefore, we approach a supervised domain adaptation scenario where the label spaces of the source and target domains are available but disjoint during training. The classifier is tasked with generalizing to the complete target domain where classes are not only from the target label space but also from the source label space. We introduce a novel CycleGAN variant, FCCGAN, which translates source images into target-stylized images that preserve their class-specific features. To further encourage the classifier to learn domain-invariant representations, we pre-train the classifier exclusively on the target domain and then employ supervised contrastive learning on source, target, and target-stylized images. We demonstrate that this framework outperforms existing domain adaptation methods in a fine-grained classification task under the disjoint label space assumption. Code and supplementary material is available at: https://github.com/enricokrohmer/sda_dls.