Learning to Adapt to Label-Scarce Image Domain via Angular Distance-Based Feature Alignment

Most recent domain adaptation (DA) methods deal with unsupervised setup, which requires numerous target images for training. However, constructing a large-scale image set of the is occasionally much harder than preparing smaller number and label pairs. To cope problem, great attention recently paid to supervised (SDA), takes an extremely small amount labeled training (e.g., at most three examples per category). In SDA adapting deep networks towards very challenging due lack data, we tackle this problem as follows. Given from source domains, first extract features project them hyper-spherical space via l2-normalization. Afterwards, additive angular margin loss embedded so that both domains are compactly grouped on basis shared class prototypes. further relieve discrepancy, pairwise spherical feature alignment incorporated. All our functions defined in space, advantage each ingredient analyzed literature. Comparative evaluation results demonstrate proposed approach superior existing methods, achieving 60.7% (1-shot) 64.4% (3-shot) average accuracies DomainNet benchmark dataset using ResNet-34 backbone. addition, by applying semi-supervised learning scheme network initialized method, achieve state-of-the-art performance (SSDA) well.