Balanced and robust unsupervised Open Set Domain Adaptation via joint adversarial alignment and unknown class isolation

Abstract

Open Set Domain Adaptation aims to implement knowledge transfer from the label-rich source domain to the label-sparse or unlabeled target domain in the scenario where the target domain contains unknown classes that are absent in the source domain. Most prevalent approaches produce pseudo labels or weights to separate the known and unknown classes in the target domain. Subsequently, these methods mitigate the domain discrepancy among known classes through adversarial alignment or by minimizing metrics that quantify the inter-domain dissimilarity. Nevertheless, the pseudo labels and weights generated by these methods can significantly deviate from the ground truth in the presence of substantial domain discrepancy, resulting in diminished accuracy when detecting unknown classes. To address this issue, we propose a balanced and robust unsupervised open set domain adaptation method which follows an end-to-end manner. Specifically, we initially extract new feature representations by a feature extractor. Subsequently, we establish the Ensemble-Binary Deviation Network to compute the deviation scores of samples in the new feature space. We maximize the divergence in deviation scores among various classes to sharpen the inter-class boundaries. Simultaneously, we employ two classifiers in the new feature space to align the feature distributions. Particularly, we initially identify the hard-to-classify samples by maximizing the discrepancy of two classifiers. Secondly, we enforce the feature extractor to minimize the discrepancy of two classifiers, aligning the feature distribution and extracting more discriminative new features. Extensive experiments on three benchmarks verify that our method achieves a balanced and robust performance, manifesting an enhancement ranging from approximately 4.9% to 7.0% over the compared methods.