Research map / Label-efficient and robust learning / Domain adaptation and generalization
Revisiting Source-Free Domain Adaptation: Insights into Representativeness, Generalization, and Variety
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Provides a theoretical upper bound on target empirical risk for source-free domain adaptation (representativeness, generalization, variety) and a progressive intersection-based sample selection + certainty-weighted semantic alignment method that beats SHOT and other SFDA baselines on Office-Home, DomainNet-126, and VisDA-C.
Problem
Source-free domain adaptation methods rely on intuitive heuristics (data generation, self-training) without theoretical grounding about what controls target-domain empirical risk when source data is unavailable.
Approach
Derives a VC-style bound on target empirical risk decomposed into representativeness (risk on confidently pseudo-labeled subset), intra-target generalization divergence between selected and unselected target data, and variety (sample-size term); operationalizes these via intersection-based (classifier top-k ∩ cosine-to-class-center) sample selection, neighbor-based pseudo-label refinement with cer
Key results
- On Office-Home (ResNet-50) the method reaches 74.1% average accuracy, versus 71.8% for SHOT.
- On DomainNet-126 the method reaches 73.2% average accuracy, about 3.6 points above the second-best GPUE (69.6%).
- On VisDA-C (ResNet-101) the method reaches 89.1% average accuracy, versus 82.9% for SHOT.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed SFDA method instead of C-SFDA | Office-Home adaptation · Office-Home | accuracy: 73.5 -> 74.1 | Helped |
| Use the proposed SFDA method instead of I-SFDA | VisDA-C adaptation · VisDA-C | accuracy: 88.4 -> 89.1 | Helped |
| Use the proposed SFDA method instead of GPUE | DomainNet126 adaptation · DomainNet126 | accuracy: 69.6 -> 73.2 | Helped |
| Add LSA and LCE to LIM (ablation) | SFDA method · Office-Home and DomainNet126, six adaptation tasks | accuracy: 67.1 -> 73.8 | Helped |
| Add LSA and LIM to LCE (ablation) | SFDA method · Office-Home and DomainNet126, six adaptation tasks | accuracy: 65.8 -> 73.8 | Helped |
| Add LCE and LIM to LSA (ablation) | SFDA method · Office-Home and DomainNet126, six adaptation tasks | accuracy: 28.2 -> 73.8 | Helped |
| Add LSA (ablation) | LCE + LIM variant · Office-Home and DomainNet126, six adaptation tasks | accuracy: 69.0 -> 73.8 | Helped |
| Add LCE (ablation) | LSA + LIM variant · Office-Home and DomainNet126, six adaptation tasks | accuracy: 68.9 -> 73.8 | Helped |
| Add LIM (ablation) | LSA + LCE variant · Office-Home and DomainNet126, six adaptation tasks | accuracy: 71.3 -> 73.8 | Helped |
| Replace C-sampling with I-sampling (ablation) | SFDA sample selection · Office-Home A→C | accuracy: 59.9 -> 61.2 | Helped |
| Replace C-sampling with I-sampling (ablation) | SFDA sample selection · DomainNet126 Re→Cl | accuracy: 76.5 -> 77.7 | Helped |
| Replace T-sampling with I-sampling (ablation) | SFDA sample selection · Office-Home A→C | accuracy: 59.1 -> 61.2 | Helped |
Surprising findings
- Using the selected target samples alone (LSA only) collapses accuracy to 28.2% on Office-Home subset tasks, showing that pseudo-label-based training without regularization degrades the source model.
Limitations
- Theoretical bound relies on assumptions that selected pseudo-labels are highly accurate and uses an empirical divergence proxy that is not directly shown to be tight.
- Gains over strong SFDA baselines on Office-Home are modest (e.g., +0.9 over SHOT+DPC on average).
Experimental setup
Datasets: Office-Home DomainNet-126 VisDA-C
Benchmarks: Office-Home DomainNet-126 VisDA-C
Metrics: top-1 accuracy average accuracy
Baselines: ERM CDAN CST SHOT A2Net AaD CoWA DaC C-SFDA I-SFDA SHOT+DPC GPUE
Builds on: SHOT Ben-David domain adaptation theory information maximization CDAN
Closest papers by meaning
- Diffusion-Driven Progressive Target Manipulation for Source-Free Domain Adaptation (NeurIPS 2025 · similarity 0.93)
- Understanding and Improving Source-free Domain Adaptation from a Theoretical Perspective (CVPR 2024 · similarity 0.92)
- Distinguish Then Exploit: Source-free Open Set Domain Adaptation via Weight Barcode Estimation and Sparse Label Assignment (CVPR 2025 · similarity 0.91)
- Dynamic Target Distribution Estimation for Source-Free Open-Set Domain Adaptation (AAAI 2025 · similarity 0.91)
- Gradient-Based Sample Selection for Black-Box Universal Domain Adaptation (AAAI 2025 · similarity 0.91)
- Proxy Denoising for Source-Free Domain Adaptation (ICLR 2025 · similarity 0.90)
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.