We call our approach TriGAN because it is based on three different factors of the images: (i) We propose the first generative MSDA method. Our main contributions can be summarized as follows. Onto a new domain-and-style specific distribution with Whitening and Coloring ( W C) batch transformations, according to the target data.Ĭontributions. Then, in the decoder, we project the intermediate invariant representation Inspired by, in the encoder we use whitening layers which progressively align the style-and-domain feature distributions. In order to modify the underlying distribution of a set of features,
Symmetrically, the decoder transforms the intermediate representations first projecting these features onto a domain-specific distribution and then onto a style-specific distribution. Our encoder obtains the intermediate representations in a two-step process: we first generate style-invariant representations and then we compute the domain-invariant representations. The content is what we want to keep unchanged during the translation process: typically, it is the foreground object shape which is described by the image labels associated with the source data samples. On the other hand, the style factor represents properties which are shared among different local parts of a single image and describes low-level features which concern a specific image (e.g., the color or the texture). The domain models properties that are shared by the elements of a dataset but which may not be shared by other datasets. To make this image translation effective, we assume that the appearance of an image depends on three factors: the content, the domain and the style. We achieve this goal using domain-invariant intermediate features, computed by the encoder part of our generator, and then projecting these features onto the domain-specific target distribution using the decoder. The translator network is “universal” because the number of parameters which need to be optimized should scale linearly with the number of domains.
In more detail, our goal is to build and train a “universal” translator which can transform an image from an input domain to a target domain. We are the first to show how it can be effectively used in a MSDA setting. While this strategy has been recently adopted in the single-source UDA scenario , Then the synthetically generated images are usedįor training the target classifier. Specifically, we generate artificial target samplesīy “translating” images from all the source domains into target-like images. In this paper we deal with (unsupervised) MSDA using a data-augmentation approach based onĪ Generative Adversarial Network (GAN). We test our approach using common MSDA benchmarks, showing that it Way, new labeled images can be generated which are used to train a final targetĬlassifier. Representation onto the pixel space using the target domain and style. The dependence from the content is kept, and then re-project this invariant This reason we propose to project the image features onto a space where only (characterized in terms of low-level features variations) and the content. The appearance of a given image depends on three factors: the domain, the style Our method is inspired by the observation that Propose the first approach for Multi-Source Domain Adaptation (MSDA) based on However, in practicalĪpplications, we typically have access to multiple sources. To the target domain from a single source dataset. Most domain adaptation methods consider the problem of transferring knowledge TriGAN: Image-to-Image Translation for Multi-Source Domain Adaptation
0 kommentar(er)
