Source-Guided Adversarial Learning and Data Augmentation for Domain Generalization
- PDF / 2,530,762 Bytes
- 12 Pages / 595.276 x 790.866 pts Page_size
- 58 Downloads / 262 Views
ORIGINAL RESEARCH
Source‑Guided Adversarial Learning and Data Augmentation for Domain Generalization Yujui Chen1 · Tse‑Wei Lin1 · Chiou‑Ting Hsu1 Received: 31 March 2020 / Accepted: 16 October 2020 © Springer Nature Singapore Pte Ltd 2020
Abstract Domain generalization aims to learn a generalized feature representation across multiple source domains so as to adapt to an unseen target domain. In this paper, we focus on image classification and propose a domain generalization framework with two cooperative ideas. First, to leverage the generalization capability, we propose a novel data augmentation method through a feature generator. The generated latent data not only preserve class-discriminative image content but also exhibit a diverse range of styles covering multiple source domains. Second, to enhance the class discriminability, we resort to the prominent adversary learning under a novel source-guided distribution constraint. We initialize the distribution prior with a mixture of Gaussians and then refer to source domains to update the constraint along with the model learning. The two models, data augmentation and adversary learning, are jointly trained to support each other and to boost the overall classification performance. Experimental results on several benchmark cross-domain datasets show that the proposed method significantly outperforms previous methods. Keywords Domain generalization · Data augmentation · Source-guided adversarial learning · Feature generator · Image classification · Gaussian mixture models
Introduction With the rapid advance of deep learning, the demand for large supervised datasets is rapidly increasing. To ease the burden of collecting large datasets, domain adaptation has attracted considerable attention. Especially, unsupervised domain adaptation aims to learn a model from “labeled” data (i.e., source domain) and then adapts the model to “unlabeled” testing data (i.e., target domain). Because there usually exists a domain shift between source and target domains, This article is part of the topical collection “Machine Learning in Pattern Analysis” guest edited by Reinhard Klette, Brendan McCane, Gabriella Sanniti di Baja, Palaiahnakote Shivakumara and Liang Wang. * Chiou‑Ting Hsu [email protected] Yujui Chen [email protected] Tse‑Wei Lin [email protected] 1
Department of Computer Science, National Tsing Hua University, Hsinchu, Taiwan
many efforts [10, 11, 29–31, 33, 34] have focused on minimizing the domain shift. For example, several methods [11, 29, 34] take advantage of adversarial learning using a domain classifier [11] to bring the source and target domains to become indistinguishable. In addition to domain adaptation, the problem of domain generalization also attracts increasing attention. Under the scenario of domain generalization, only multiple source domains are given but no target domain is available during the training stage. Because target domain data are totally “invisible”, there is no way to estimate the target distribution or to minimize the
Data Loading...
