Research map / Label-efficient and robust learning / Domain adaptation and generalization
UDA-Bench: Revisiting Common Assumptions in Unsupervised Domain Adaptation Using a Standardized Framework
Tarun Kalluri, Sreyas Ravichandran, Manmohan Chandraker
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 |
|---|---|---|---|
| use Swin-V2-t instead of ResNet-50 backbone for source-only transfer (ablation) | Source-only baseline · DomainNet (R->C) | relative drop sigma_st: 45.21 -> 35.45 | Improved |
| use Swin backbone with plain source-only training instead of UDA with ResNet-50 | Source-only baseline · DomainNet (R->C) | target accuracy: 54.5 (SOTA UDA, ResNet-50) vs 55.5 (Swin source-only) | Improved |
| apply UDA methods on stronger backbones (Swin, DeiT) instead of ResNet-50 (ablation) | UDA methods · DomainNet (R->C) | relative gain over source-only: 20% (ResNet-50) -> 7% (Swin), 10% (DeiT) | Hurt |
| apply UDA methods on Swin instead of ResNet-50 (ablation) | UDA methods · OfficeHome (Ar->Pr) | relative gain over source-only: 18% -> 8% | Hurt |
| reduce unlabeled target data from 100% to 25% (ablation) | UDA methods · DomainNet and VisDA | target accuracy: <1% drop | No clear effect |
| double source labeled data from 50% to 100% (ablation) | UDA methods · DomainNet (R->C) | target accuracy: ~9% average gain vs <0.5% for doubling target unlabeled data | Improved |
| pre-train supervised on iNaturalist 1M instead of ImageNet 1M for birds adaptation (ablation) | Plain source-only transfer · CUB->Drawing | target accuracy: 50.2 -> 58.77 | Improved |
| pre-train supervised on Places-205 1M instead of ImageNet 1M for scene adaptation (ablation) | Plain source-only transfer · GeoPlaces USA->Asia | target accuracy: 34.55 -> 41.95 | Improved |
| use mismatched pre-training dataset (e.g. Places for birds) (ablation) | UDA methods · CUB->Drawing and DomainNet | target accuracy: >10% drop in most cases | Hurt |
| add in-task pre-training on top of UDA methods (MemSAC, ToAlign) (ablation) | UDA methods · CUB200 | target accuracy: 17.1% and 17.3% gains over ImageNet pre-training source-only | Improved |
| use object-centric ImageNet pre-training instead of scene-centric Places for self-supervised (ablation) | SwAV, MoCo-V3, MAE · DomainNet and CUB200 | target accuracy: MAE DomainNet plain 34.76 -> 38.58 | Improved |
| apply SAFN and MCD UDA methods relative to source-only baseline | SAFN, MCD · Additional UDA methods | relative gain over source-only | Hurt |
Closest papers by meaning
- Can We Evaluate Domain Adaptation Models Without Target-Domain Labels? (ICLR 2024 · similarity 0.93)
- Prior Knowledge Guided Unsupervised Domain Adaptation (ECCV 2022 · similarity 0.93)
- Make the U in UDA Matter: Invariant Consistency Learning for Unsupervised Domain Adaptation (NeurIPS 2023 · similarity 0.92)
- Towards Reliable Model Selection for Unsupervised Domain Adaptation: An Empirical Study and A Certified Baseline (NeurIPS 2024 · similarity 0.92)
- Unsupervised Multi-Source Domain Adaptation Without Access to Source Data (CVPR 2021 · similarity 0.91)
- Tackling Dimensional Collapse toward Comprehensive Universal Domain Adaptation (ICML 2025 · similarity 0.89)
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.
