Search

Unsupervised domain adaptation aims to transfer knowledge from labeled source domain to unlabeled target domain. Traditional methods usually achieve domain adaptation by aligning the distributions between two domains once. We propose to align the distributions twice by a disentanglement and reconstruction process, named DTR (Disentanglement Then Reconstruction). Specifically, a feature extraction network shared by both source and target domains is used to obtain the original extracted features, then the domain invariant features and domain specific features are disentangled from the original extracted features. The domain distributions are explicitly aligned when disentangling domain invariant features. Based on the disentangled features, the class prototypes and domain prototypes can be estimated. Then, a reconstructor is trained by the disentangled features. By this reconstructor, we can construct prototypes in the original feature space using the corresponding class prototype and domain prototype similarly. The extracted features are forced to close the corresponding constructed prototypes. In this process, the distribution between two domains is implicitly aligned again. Experiment results on several public datasets confirm the effectiveness of our method. • Our method can use domain invariant features and domain specific features very well. • By learning more compact features, domain distributions will be aligned again. • The experiments on public datasets are conducted and the proposed method works well. • Thorough analysis experiments also illustrates the advantages of our method. [ABSTRACT FROM AUTHOR]

Source-free domain adaptation requires no access to the source domain training data during unsupervised domain adaption. This is critical for meeting particular data sharing, privacy, and license constraints, whilst raising novel algorithmic challenges. Existing source-free domain adaptation methods rely on either generating pseudo samples/prototypes of source or target domain style, or simply leveraging pseudo-labels (self-training). They suffer from low-quality generated samples/prototypes or noisy pseudo-label target samples. In this work, we address both limitations by introducing a novel Class Prototype Discovery (CPD) method. In contrast to all alternatives, our CPD is established on a set of semantic class prototypes each constructed for representing a specific class. By designing a classification score based prototype learning mechanism, we reformulate the source-free domain adaptation problem to class prototype optimization using all the target domain training data, and without the need for data generation. Then, class prototypes are used to cluster target features to assign them pseudo-labels, which highly complements the conventional self-training strategy. Besides, a prototype regularization is introduced for exploiting well-established distribution alignment based on pseudo-labeled target samples and class prototypes. Along with theoretical analysis, we conduct extensive experiments on three standard benchmarks to validate the performance advantages of our CPD over the state-of-the-art models. • We propose a novel Class Prototype Discovery method for solving the SFDA problem. • A prototype regularization is introduced based on distribution alignment strategy. • CPD outperforms a wide variety of state-of-the-art methods, often by a large margin. [ABSTRACT FROM AUTHOR]