Multi-Component Image Translation for Deep Domain Generalization

Abstract

Domain adaption (DA) and domain generalization (DG) are two closely related methods which are both concerned with the task of assigning labels to an unlabeled data set. The only dissimilarity between these approaches is that DA can access the target data during the training phase, while the target data is totally unseen during the training phase in DG. The task of DG is challenging as we have no earlier knowledge of the target samples. If DA methods are applied directly to DG by a simple exclusion of the target data from training, poor performance will result for a given task. In this paper, we tackle the domain generalization challenge in two ways. In our first approach, we propose a novel deep domain generalization architecture utilizing synthetic data generated by a Generative Adversarial Network (GAN). The discrepancy between the generated images and synthetic images is minimized using existing domain discrepancy metrics such as maximum mean discrepancy or correlation alignment. In our second approach, we introduce a protocol for applying DA methods to a DG scenario by excluding the target data from the training phase, splitting the source data to training and validation parts, and treating the validation data as target data for DA. We conduct extensive experiments on four cross-domain benchmark datasets. Experimental results signify our proposed model outperforms the current state-of-the-art methods for DG.

Correlation-aware adversarial domain adaptation and generalization

“…For domain generalization, the training data always contains more than one source domain. Most of the existing domain generalization methods [39,20,22,38] split the source data as 70% -30%…”

Semantic Consistency and Identity Mapping Multi-Component Generative Adversarial Network for Person Re-Identification

“…Generally, GANs use noise to synthesise an image and the network is trained in an adversarial manner. Inspired by the success of GANs, various extensions have been proposed for image-to-image translation [13,28], pixel-level transfer from source to target domains [43], and style transfer between domains [42]. Rather than using noise alone as the stimulus, the conditional GAN (cGAN) [17] is proposed to control the mode of generated images, however, cGANs need a pair of images for training which is not available for many tasks.…”

Source-Guided Adversarial Learning and Data Augmentation for Domain Generalization

“…Because target domain data are totally "invisible", there is no way to estimate the target distribution or to minimize the domain shift when developing the model. Therefore, most domain generalization methods [3,7,13,21,25,27,32] rely on learning a domain-invariant representation from multiple source domains so as to adapt or to generalize to the unseen target domain. Some methods [14,27] rely on image translation techniques by transferring images from one domain to the other and then use the transferred images as augmented training data to improve the model generalization.…”

“…Therefore, most domain generalization methods [3,7,13,21,25,27,32] rely on learning a domain-invariant representation from multiple source domains so as to adapt or to generalize to the unseen target domain. Some methods [14,27] rely on image translation techniques by transferring images from one domain to the other and then use the transferred images as augmented training data to improve the model generalization. Other methods [13,21,22,27] focus on learning a generalized representation by minimizing discrepancy among source domains.…”

“…Some methods [14,27] rely on image translation techniques by transferring images from one domain to the other and then use the transferred images as augmented training data to improve the model generalization. Other methods [13,21,22,27] focus on learning a generalized representation by minimizing discrepancy among source domains. For example, in [13], Ghifary et al proposed a multi-task autoencoder with separate branches to relate to different source domains.…”