One-Sided Unsupervised Domain Mapping

By Sagie Benaim, Lior Wolf. Neural Information Processing Systems, 2017

334 citations recorded by the source corpus.

In unsupervised domain mapping, the learner is given two unmatched datasets $A$ and $B$. The goal is to learn a mapping $G_{AB}$ that translates a sample in $A$ to the analog sample in $B$. Recent approaches have shown that when learning simultaneously both $G_{AB}$ and the inverse mapping $G_{BA}$, convincing mappings are obtained. In this work, we present a method of learning $G_{AB}$ without learning $G_{BA}$. This is done by learning a mapping that maintains the distance between a pair of samples. Moreover, good mappings are obtained, even by maintaining the distance between different parts of the same sample before and after mapping. We present experimental results that the new method not only allows for one sided mapping learning, but also leads to preferable numerical results over the existing circularity-based constraint. Our entire code is made publicly available at this https URL .

Subfields: Artificial Intelligence, Media Technology

View Paper: https://www.semanticscholar.org/paper/fd4537b92ab9fa7c653e9e5b9c4f815914a498c0

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Media Technology

Citations in the corpus

Alleviating Semantics Distortion in Unsupervised Low-Level Image-to-Image Translation via Structure Consistency Constraint (2022, background): “However, since paired data are often unavailable or expensive to obtain, unsupervised I2I translation has attracted intense attention in recent years [3, 17, 25, 26, 32, 43, 69, 73].”
An Introduction to Image Synthesis with Generative Adversarial Nets (2018, methodology): “One interesting thing is, as stated in [79], that Distance-GAN computes the distances in raw RGB space and still achieves better performance than baselines, but it may help if the distances are calculated in images’ latent feature space where the features can be extracted using pre-trained image…”
Image-Image Domain Adaptation with Preserved Self-Similarity and Domain-Dissimilarity for Person Re-identification (2017, background): “In [3], an unsupervised distance loss is proposed for one side domain mapping.”
Unpaired Image-to-Image Translation via Latent Energy Transport (2020, background): “, DistanceGAN [1], GcGAN [10] and CUT [31].”
Zero-shot Domain Adaptation without Domain Semantic Descriptors (2018, background): “Unsupervised domain adaptation aims to adapt a target domain by using labeled source data and unlabeled target data [37, 23, 22, 5, 3].”
Multimodal Unsupervised Image-to-Image Translation (2018, background): “Some works enforce the translation to preserve certain properties of the source domain data, such as pixel values [21], pixel gradients [22], semantic features [10], class labels [22], or pairwise sample distances [16].”
Unsupervised Multi-Target Domain Adaptation: An Information Theoretic Approach (2018, background): “Recently proposed unsupervised DA methods [27]–[30] operate by training deep neural networks using adversarial training, which allows the learning of feature representations that are simultaneously discriminative of source labels, and indistinguishable between the source and target domain.”
Few-Shot Unsupervised Image-to-Image Translation (2019, background): “For example, some works enforce the translation to preserve certain properties of the source data, such as pixel values [38], pixel gradients [5], semantic features [43], class labels [5], or pairwise sample distances [3].”
Unpaired Image-to-Speech Synthesis With Multimodal Information Bottleneck (2019, background): “Since then, numerous methods have tackled the problem from an unsupervised learning perspective, eliminating the need for paired data [58, 41, 26, 5, 21].”
Image Manipulation with Perceptual Discriminators (2018, background): “datasets for training [41,19], recently some impressive results have been obtained using unaligned training data, where only empirical distributions of the input and the output images are provided [42,5,43].”
One-Shot Unsupervised Cross Domain Translation (2018, methodology): “We have made attempts to compare various image parts within x, thereby generalizing the imagehalves solution of [5].”
Contrastive Learning for Unpaired Image-to-Image Translation (2020, methodology): “85 Distance [4] 8.”
EEG-Based Emotion Recognition Using Regularized Graph Neural Networks (2019, methodology): “The domain mapping technique is typically applied in the computer vision field where pixel-level image-to-image translation from one domain to another domain improves domain adaptation performance [49].”
Few-shot Image Generation via Cross-domain Correspondence (2021, background): “To alleviate mode collapse in GANs, DistanceGAN [2] proposes to preserve the distances between input pairs in the corresponding generated output pairs.”
InstaGAN: Instance-aware Image-to-Image Translation (2018, methodology): “However, we remark that training two coupled mappings is not essential for our method, and one can also design a single mapping following other approaches (Benaim & Wolf, 2017; Galanti et al., 2018).”
Feedback Adversarial Learning: Spatial Feedback for Improving Generative Adversarial Networks (2019, neutral): “With the difficulty of collecting interesting paired data, many works [68, 35, 60, 34, 6, 48] have proposed alternative methods to translate images.”
How Generative Adversarial Networks and Its Variants Work : An Overview of GAN (2017, methodology): “[35]: Standard GAN, [73]: LSGAN, [138]: EBGAN, [6]: WGAN, [97]: Improved WGAN, [80]: McGAN, [10]: CramerGAN, [65]: MMDGAN, [79]: FisherGAN, [76]: Unrolled GAN, [14]: MRGAN, [59]: DRAGAN, [77]: Conditional GAN, [16]: InfoGAN, [106]: ss-InfoGAN, [27]: ALI, [25]: BiGAN, [116]: AGE, [58]: ARAE, [61]: VA”