Research map / Label-efficient and robust learning / Domain adaptation and generalization
SF(DA)$^2$: Source-free Domain Adaptation Through the Lens of Data Augmentation
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.
Proposes SF(DA)^2, a source-free domain adaptation method that builds a feature-space augmentation graph, clusters target predictions spectrally (SNC), and adds closed-form implicit feature augmentation (IFA) and feature disentanglement (FD) regularizers.
Problem
Source-free domain adaptation needs to adapt pretrained models to unlabeled target domains, but explicit data augmentation for SFDA requires prior knowledge of class-preserving transformations and adds memory and compute cost.
Approach
Define an augmentation graph over target features using feature-space nearest neighbors as positive pairs (no explicit augmentation), partition it spectrally in the prediction space, and approximate an infinite number of Gaussian-augmented features with a closed-form IFA upper bound, regularized by a covariance-based feature disentanglement loss.
Key results
- On VisDA (ResNet-101), SF(DA)^2 reaches 88.1% per-class accuracy, above AaD (87.3%) and the best other source-free method listed.
- On DomainNet-126 (7 shifts, ResNet-50), SF(DA)^2 averages 68.3% versus 67.1% for SHOT.
- On VisDA-RSUT (highly imbalanced), SF(DA)^2 achieves 77.3% per-class accuracy versus 73.4% for AaD.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SF(DA)^2 instead of the strongest reported SFDA baseline | VisDA adaptation · VisDA | Per-class accuracy (%): 87.3 -> 88.1 | Helped |
| Use SF(DA)^2 instead of the strongest reported SFDA baseline | DomainNet-126 adaptation · DomainNet-126, 7 domain shifts | Avg. accuracy (%): 67.1 -> 68.3 | Helped |
| Use SF(DA)^2 instead of the strongest reported SFDA baseline | PointDA-10 adaptation · PointDA-10 | Avg. accuracy (%): 55.9 -> 57.1 | Helped |
| Use SF(DA)^2 instead of the strongest reported SFDA baseline | VisDA-RSUT adaptation · VisDA-RSUT | Per-class accuracy (%): 73.4 -> 77.3 | Helped |
| Add IFA and FD losses | CoWA-JMDS and SNC · VisDA-RSUT | Performance: +3%p | Helped |
| Add IFA loss alone to SNC (ablation) | SNC adaptation · VisDA-RSUT | Average accuracy | Hurt |
| Add FD loss alone to SNC (ablation) | SNC adaptation · VisDA-RSUT | Average accuracy | Hurt |
| Add IFA and FD losses jointly to SNC (ablation) | SNC adaptation · VisDA-RSUT | Average accuracy | Helped |
| Add IFA loss | SHOT · SFDA settings reported in Figure (b) | Performance | Hurt |
| Vary IFA and FD loss hyperparameters α1 and α2 (ablation) | SF(DA)^2 · VisDA | Performance | No clear effect |
Surprising findings
- Adding IFA loss alone to SNC decreases performance; only the combination with FD yields gains.
- IFA slightly decreases SHOT's performance, since SHOT's assumptions do not align well with IFA.
Limitations
- Assumes closed-set setting where target classes match source classes.
- Adding IFA alone can degrade performance when the feature space is entangled; requires FD to help.
- Relies on the clustering assumption of the source model and neighbor relationships in feature space.
Experimental setup
Datasets: VisDA DomainNet PointDA-10 VisDA-RSUT Office-31
Benchmarks: VisDA DomainNet-126 PointDA-10 VisDA-RSUT Office-31
Metrics: per-class accuracy average accuracy harmonic mean of accuracy F1-score
Baselines: SHOT NRC AaD CoWA-JMDS 3C-GAN MCC TENT VDM-DA PointDAN SHOT++
Builds on: spectral contrastive loss ISDA SHOT NRC AaD
Closest papers by meaning
- Understanding and Improving Source-free Domain Adaptation from a Theoretical Perspective (CVPR 2024 · similarity 0.89)
- Revisiting Source-Free Domain Adaptation: Insights into Representativeness, Generalization, and Variety (CVPR 2025 · similarity 0.89)
- Open-Set Heterogeneous Domain Adaptation: Theoretical Analysis and Algorithm (AAAI 2025 · similarity 0.87)
- Source-Free Domain Adaptation with Frozen Multimodal Foundation Model (CVPR 2024 · similarity 0.87)
- Diffusion-Driven Progressive Target Manipulation for Source-Free Domain Adaptation (NeurIPS 2025 · similarity 0.87)
- Dynamic Target Distribution Estimation for Source-Free Open-Set Domain Adaptation (AAAI 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.