Unsupervised Open Domain Recognition by Semantic Discrepancy Minimization

Abstract

We address the unsupervised open domain recognition (UODR) problem, where categories in labeled source domain S is only a subset of those in unlabeled target domain T . The task is to correctly classify all samples in T including known and unknown categories. UODR is challenging due to the domain discrepancy, which becomes even harder to bridge when a large number of unknown categories exist in T . Moreover, the classification rules propagated by graph CNN (GCN) may be distracted by unknown categories and lack generalization capability.To measure the domain discrepancy for asymmetric label space between S and T , we propose Semantic-Guided Matching Discrepancy (SGMD), which first employs instance matching between S and T , and then the discrepancy is measured by a weighted feature distance between matched instances. We further design a limited balance constraint to achieve a more balanced classification output on known and unknown categories. We develop Unsupervised Open Domain Transfer Network (UODTN), which learns both the backbone classification network and GCN jointly by reducing the SGMD, enforcing the limited balance constraint and minimizing the classification loss on S. UODTN better preserves the semantic structure and enforces the consistency between the learned domain invariant visual features and the semantic embeddings. Experimental results show superiority of our method on recognizing images of both known and unknown categories.

Fast Batch Nuclear-norm Maximization and Minimization for Robust Domain Adaptation

“…We apply our method to three domain adaptation tasks, i.e., unsupervised domain adaptation (UDA), semi-supervised domain adaptation (SSDA), and unsupervised open domain recognition (UODR). The experiments of the three tasks are accomplished on datasets of Office-31 [18], Office-Home [51], Balanced Domainnet, Semi Domainnet [22] and I2AwA [15]. We also denote the direct entropy minimization and batch Frobenius-norm maximization as EntMin and BFM in our experiments.…”

“…Visual domain adaptation [18] has gained significant improvement in the past few years. Recently, the study of DA methods focuses on diverse adaptation circumstances, such as unsupervised DA [4], [13], [19], multi-source DA [20], [21], semi-supervised DA [22] and unsupervised open DA [15]. Similar to existing methods of [16], [19], [23], we explore the transfer properties in DA, to construct general frameworks suitable for various DA scenarios.…”

“…To maintain prediction diversity, a direct technique is imbalanced learning [38]. Typical methods such as [15], [39] directly enforce the ratio of predictions on minority categories. However, the category distribution is required as prior knowledge.…”

“…Nevertheless, directly applying Shannon entropy minimization also leads to the reduction of diversity [13]. To encourage the diversity, balance constraints [14], [15] is developed to equilibrate the output distribution among categories under the guidance of various types of distribution priors. When the distribution priors are unavailable, the predicted pseudo-labels are generated to approximate the statistical distribution [9], [10] to encourage higher prediction probability on the minority categories, which seems less straightforward towards more accurate prediction model.…”

Domain Conditioned Adaptation Network

“…Domain adaptation (DA) (Pan and Yang 2010) has been witnessed as a promising technique to transfer knowledge from a well-labeled and related source domain to assist the target learning (Baktashmotlagh et al 2013;Donahue et al 2014;Sun and Saenko 2016;Zhuo et al 2019). Previous DA methods generally aim to align domain distributions by either reweighting instances (Chen, Weinberger, and Blitzer 2011) or learning domain-invariant features (Ghifary et al 2017).…”