Research map / Label-efficient and robust learning / Domain adaptation and generalization
Deep Domain-Adversarial Image Generation for Domain Generalisation
Kaiyang Zhou, Yongxin Yang, Timothy Hospedales, Tao Xiang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| augment source training examples with adversarially generated unseen-domain images | DDAIG · Digits-DG Avg. | accuracy: 73.7 -> 77.6 | Improved |
| learn a CNN transformation network rather than use gradient-based perturbations | DDAIG vs CrossGrad · PACS Avg. | accuracy: 80.7 -> 83.1 | Improved |
| augment source data with generated unseen-domain images | DDAIG · Office-Home Avg. | accuracy: 64.7 -> 65.5 | Improved |
| augment source training data with generated images | DDAIG · Market1501→Duke | R1: 48.5 -> 50.6 | Improved |
| augment source training data with generated images | DDAIG · Duke→Market1501 | R1: 57.7 -> 60.9 | Improved |
| vary transformation strength λ from 0.1 to 0.7 (ablation) | DDAIG · Digits-DG | accuracy: 76.8; 76.8; 76.6; 76.8 | No clear effect |
| train with generated novel data only, removing source examples (ablation) | DDAIG · Digits-DG Avg. | accuracy: 73.7 -> 72.9 | Hurt |
| combine generated novel data with source data instead of using source data alone (ablation) | DDAIG · Digits-DG Avg. | accuracy: 73.7 -> 77.6 | Improved |
| add an STN before the perturbation network (ablation) | DDAIG · Rotated MNIST 40° | accuracy: 54.1 -> 61.1 | Improved |
| add an STN before the perturbation network (ablation) | DDAIG · Rotated MNIST 50° | accuracy: 40.3 -> 46.6 | Improved |
| use learned CNN perturbations instead of CrossGrad perturbations | DDAIG · Rotated MNIST 40° | accuracy: 45.6 -> 54.1 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- MetaReg — MetaReg: Towards Domain Generalization using Meta-Regularization (NeurIPS 2018) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Deep Domain Generalization via Conditional Invariant Adversarial Networks (ECCV 2018 · similarity 0.89)
- Deeper, Broader and Artier Domain Generalization (ICCV 2017 · similarity 0.88)
- Adversarial Teacher-Student Representation Learning for Domain Generalization (NeurIPS 2021 · similarity 0.88)
- Episodic Training for Domain Generalization (ICCV 2019 · similarity 0.88)
- Domain Generalization by Learning and Removing Domain-specific Features (NeurIPS 2022 · similarity 0.88)
- Adversarial Domain Prompt Tuning and Generation for Single Domain Generalization (CVPR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
