Domain Adaptive Few-Shot Open-Set Learning

Few-shot learning has made impressive strides in addressing the crucialchallenges of recognizing unknown samples from novel classes in target querysets and managing visual shifts between domains. However, existing techniquesfall short when it comes to identifying target outliers under domain shifts bylearning to reject pseudo-outliers from the source domain, resulting in anincomplete solution to both problems. To address these challengescomprehensively, we propose a novel approach called Domain Adaptive Few-ShotOpen Set Recognition (DA-FSOS) and introduce a meta-learning-based architecturenamed DAFOSNET. During training, our model learns a shared and discriminativeembedding space while creating a pseudo open-space decision boundary, given afully-supervised source domain and a label-disjoint few-shot target domain. Toenhance data density, we use a pair of conditional adversarial networks withtunable noise variances to augment both domains closed and pseudo-open spaces.Furthermore, we propose a domain-specific batch-normalized class prototypesalignment strategy to align both domains globally while ensuringclass-discriminativeness through novel metric objectives. Our training approachensures that DAFOS-NET can generalize well to new scenarios in the targetdomain. We present three benchmarks for DA-FSOS based on the Office-Home,mini-ImageNet/CUB, and DomainNet datasets and demonstrate the efficacy ofDAFOS-NET through extensive experimentation