[論文レビュー] One-Sided Unsupervised Domain Mapping

Sagie Benaim, Lior Wolf|arXiv (Cornell University)|Jun 2, 2017

Domain Adaptation and Few-Shot Learning参考文献 18被引用数 55

ひとこと要約

この論文は、一方的な教師なしドメインマッピング手法（DistanceGAN）を提示し、逆写像を必要とせずにドメインAからドメインBへのマッピングを学習するためにクロスドメインの組み合わせ距離を保持し、複数のデータセットで循環性ベースの制約より性能が向上することを示します。

ABSTRACT

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 https://github.com/sagiebenaim/DistanceGAN .

研究の動機と目的

• Motivate and enable learning a mapping from domain A to domain B without paired samples or learning the inverse mapping.
• Propose a distance-based constraint that enforces correlation between pairwise distances before and after mapping.
• Show that distance-based one-sided mapping yields competitive or superior results compared to circularity-based constraints across multiple datasets.
• Analyze the practical implications of distance-based constraints, including self-distance constraints, for stable one-sided translation.

提案手法

• Introduce a distance-based loss that maximizes correlation between pairwise distances in A and the mapped space B.
• Replace or augment circularity constraints with a distance preservation objective: Ldistance = E|d_k − d′_k| with normalized distances in A and B.
• Propose a self-distance constraint that compares left/right halves of the same image to encourage stable mappings.
• Train using adversarial losses for GAB (A→B) and GBA (B→A) in combination with distance-based and self-distance losses, using architectures from DiscoGAN and CycleGAN.
• Demonstrate that learning GAB without GBA is feasible and can yield better numerical results than circularity baselines; provide per-dataset parameter settings in experiments.

実験結果

リサーチクエスチョン

• RQ1 Can a one-sided mapping from domain A to domain B be learned effectively without the inverse mapping?
• RQ2 Does enforcing distance-based consistency between input and mapped outputs yield better or comparable results to cycle-consistency constraints?
• RQ3 How do self-distance constraints influence the quality and stability of one-sided domain mappings?
• RQ4 Are distance-based constraints robust across different domains and architectures (DiscoGAN/CycleGAN backbones)?

主な発見

• A one-sided distance-based constraint can successfully learn mappings from A to B without learning GBA.
• The distance-based approach often yields preferable numerical results over the circularity-based constraint across multiple datasets (e.g., CelebA, horses↔zebras, car↔head).
• Distance constraints can be applied to full RGB pixel values as well as to within-image halves, enabling flexible implementation.
• Self-distance constraints offer an additional one-sample-per-step regularization that can improve or complement distance-based methods.
• Combining distance-based with cycle constraints sometimes provides the best results, though in some datasets this combination may underperform compared to distance alone.