Source-linked AI summary
Adversarial Domain Adaptation with Domain Mixup
Minghao Xu, Jian Zhang, Bingbing Ni, Teng Li, Chengjie Wang, Qi Tian, Wenjun Zhang
TL;DR
Unsupervised domain adaptation is challenged by limited latent-space coverage and hard-label domain discrimination. DM-ADA applies pixel- and feature-level domain mixup with soft labels and reports superior performance across tasks with varied domain shift and data complexity. Its setting assumes labeled source and unlabeled target samples sharing the same label space.
Problem
Adversarial adaptation may leave much of latent space insufficiently domain-invariant and may use hard labels rather than soft scores to represent inter-domain differences.
Method
DM-ADA jointly uses pixel- and feature-level mixup, soft domain labels, flexible-margin triplet loss, and adversarial alignment across intermediate domain representations.
Results
DM-ADA achieves superior results across adaptation tasks with different domain-shift extents and data complexity, reaching state-of-the-art performance in most settings.
Takeaways & Limitations
Domain mixup promotes a more continuous domain-invariant latent distribution and gives the discriminator refined scores relative to the two domains.
Takeaways & Limitations
The method assumes labeled source samples, unlabeled target samples, and a shared label space across both domains.
Abstract
from arXiv · showhide
Recent works on domain adaptation reveal the effectiveness of adversarial learning on filling the discrepancy between source and target domains. However, two common limitations exist in current adversarial-learning-based methods. First, samples from two domains alone are not sufficient to ensure domain-invariance at most part of latent space. Second, the domain discriminator involved in these methods can only judge real or fake with the guidance of hard label, while it is more reasonable to use soft scores to evaluate the generated images or features, i.e., to fully utilize the inter-domain information. In this paper, we present adversarial domain adaptation with domain mixup (DM-ADA), which guarantees domain-invariance in a more continuous latent space and guides the domain discriminator in judging samples' difference relative to source and target domains. Domain mixup is jointly conducted on pixel and feature level to improve the robustness of models. Extensive experiments prove that the proposed approach can achieve superior performance on tasks with various degrees of domain shift and data complexity.
Introduction
Domain adaptation addresses performance loss caused by domain shift when labeled target data are unavailable. DM-ADA uses pixel- and feature-level domain mixup with soft domain labels to enrich latent representations and guide adversarial alignment.
- Domain shift can reduce performance when models trained on one domain transfer to another, motivating adaptation with unlabeled target data.
- Adversarial adaptation seeks domain-invariant features, but conventional discriminators rely on hard labels and limited source-target patterns.
- DM-ADA jointly trains a classification network and generates auxiliary source-like images from embeddings learned across both domains.
- Pixel-level mixup supplies mixed images and soft labels, while feature-level interpolation expands the latent search range and creates a more continuous domain-invariant distribution.
- Experiments on three benchmarks with varied domain shift and data complexity achieve state-of-the-art performance in most settings.
Related Work
Domain adaptation methods align domains through distribution measures, adversarial training, and pixel-level generation. DM-ADA extends these approaches by using domain mixup to improve robustness and address mismatches between generated data and hard labels.
- Domain adaptation methods include distribution alignment through MMD variants and class-conditional similarity measures.
- Adversarial methods transfer domain information by training representations or embeddings to confuse a domain discriminator.
- Other approaches model multimode structures, task-specific decision boundaries, attention regions, and structure-aware alignment.
- Pixel-level adversarial methods adapt source images toward the target domain, expand training data, or jointly generate images and train classifiers.
- DM-ADA addresses possible mismatches between generated data and hard labels by applying domain mixup at pixel and feature levels.
Adversarial Domain Adaptation
DM-ADA mixes source and target data at pixel and feature levels, then jointly performs classification, generation, and adversarial domain alignment. The framework uses soft domain labels, triplet constraints, and class-consistency objectives to promote continuous, category-level domain invariance.
- Problem: Unsupervised adaptation assumes labeled source samples and unlabeled target samples share the same K-class label space.Source and target samples follow distributions Ps and Pt, respectively.
- Domain Mixup: A shared encoder maps source and target inputs to latent mean and standard-deviation embeddings, which are linearly mixed using the same ratio to create intermediate feature representations.The encoder is regularized toward a standard Gaussian prior.
- Adversarial Domain Alignment: The training framework splits into object classification and generative adversarial branches, with the decoder generating auxiliary images from latent codes and class-related inputs.The classifier is trained with source-domain cross-entropy, while the decoder and discriminator participate in a min-max game.
- Domain Mixup: Pixel-level domain mixup linearly combines source and target images and assigns the mixture a soft domain label equal to its mixup ratio.The ratio λ is sampled from Beta(α, α), with α set to 2.0 in experiments; source and target domain labels are 1 and 0.
- Adversarial Domain Alignment: Adversarial alignment constrains source, target, and intermediate representations, using soft domain scores and a triplet loss with flexible margin f_tri(λ) = |2λ − 1|.The triplet ordering changes according to whether λ is at least 0.5, reflecting which domain contributes more to the mixup sample.
- Adversarial Domain Alignment: Feature-level mixup expands the continuous domain-invariant latent space, helping same-class cross-domain codes form clusters and refining the decision boundary.The procedure also filters low-confidence target pseudo-labels using an adaptively adjusted threshold τ.
Experiments
Experiments across digits, Office-31, and VisDA-2017 evaluate DM-ADA under varied domain shifts, while ablations examine its mixup, loss, discriminator, and label components. The reported results show strong target-domain classification, improved domain alignment, and limited sensitivity to selected hyperparameters.
- Classification on digits: DM-ADA achieves state-of-the-art performance on all four digit-transfer settings.Results are averaged over five independent runs with random initialization.
- Classification on Office-31: DM-ADA obtains the best performance in three of four hard Office-31 cases and exceeds 99.5% accuracy on the two easier cases.The evaluated tasks include transfers among Amazon, Webcam, and DSLR domains.
- Classification on VisDA-2017: DM-ADA achieves the highest accuracy among adaptation approaches on VisDA-2017 and exceeds the pretrained ResNet-101 baseline by a great margin.The task transfers from synthetic 3D-object renderings to photo-realistic images.
- Ablation study: Feature-level mixup improves over the two-domain baseline, while pixel-level mixup further improves performance by guiding discriminator scores between 0 and 1.The combined mixup design makes the latent domain-invariant space more continuous and supplies soft domain guidance.
- Ablation study: Adding triplet loss improves performance and produces the clearest class separation in target-domain t-SNE features when combined with two-level domain mixup.The cited visualization concerns the SV →MN transfer task.
- Ablation study: Adding Dcls increases classification accuracy by 1.7%, while pseudo target labels make the model’s performance state-of-the-art.Dcls helps generated images preserve input class information and supports category-aware alignment.
Conclusion
The paper addresses unsupervised domain adaptation by transferring knowledge from a source domain to a target domain with a GAN-based architecture. Domain mixup at pixel and feature levels promotes a more continuous domain-invariant latent space and supports strong performance across adaptation tasks with varying domain shift and data complexity.
- A GAN-based architecture transfers knowledge from the source domain to the target domain.
- Domain mixup at pixel and feature levels is proposed to facilitate a more continuous domain-invariant latent space and use inter-domain information.
- Extensive experiments demonstrate predominant performance across adaptation tasks with different degrees of domain shift and data complexity.