Research map / Label-efficient and robust learning / Domain adaptation and generalization
Rethinking the Role of Pre-Trained Networks in Source-Free Domain Adaptation
Wenyu Zhang, Li Shen, Chuan-Sheng Foo
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 |
|---|---|---|---|
| add co-learning with Swin-B pre-trained branch | source-trained ResNet-50 · Office-Home | avg accuracy: 59.6 -> 83.5 | Improved |
| add co-learning with ResNet-50 ImageNet branch to SHOT++ | SHOT++ · Office-31 | avg accuracy: 89.2 -> 89.1 | No clear effect |
| add co-learning with Swin-B to NRC | NRC · Office-Home | avg accuracy: 72.0 -> 81.4 | Improved |
| add co-learning with Swin-B to SHOT | SHOT · VisDA-C | avg accuracy: 82.4 -> 85.2 | Improved |
| use MatchOrConf instead of SelfConf pseudolabeling (ablation) | co-learning · Office-Home (A->C,A->P,A->R) | avg accuracy: 67.0 -> 77.7 | Improved |
| use MatchOrConf instead of MatchAndConf (ablation) | co-learning · Office-Home (A->C,A->P,A->R) | avg accuracy: 75.4 -> 77.7 | Improved |
| lower confidence threshold gamma from 0.9 to 0.1 (ablation) | co-learning · VisDA-C | accuracy: 85.5 -> 87.1 | Improved |
| raise confidence threshold gamma from 0.1 to 0.7 (ablation) | co-learning · Office-Home | accuracy: 76.1 -> 77.5 | Improved |
| finetune only the classifier in pre-trained branch vs finetuning feature extractor (ablation) | co-learning · Office-Home (A->C,A->P,A->R) | avg accuracy: 76.2 -> 77.7 | Improved |
| estimate class centroids from all samples rather than pseudolabeled only (ablation) | co-learning · Office-Home (A->C,A->P,A->R) | avg accuracy: 77.4 -> 77.7 | Improved |
| iteratively refine class centroids (3 iterations) (ablation) | co-learning · Office-Home (A->C,A->P,A->R) | avg accuracy: 77.7 -> 76.7 | Hurt |
| adapt ResNet-50 with Swin-B co-learning vs fixed Swin-B with source classification head (ablation) | Swin-B · Office-Home (A->C,A->P,A->R) | avg accuracy: 81.0 -> 83.5 | Improved |
Closest papers by meaning
- Source-Free Domain Adaptation Guided by Vision and Vision-Language Pre-training (IJCV 2024 · similarity 0.93)
- Understanding and Improving Source-free Domain Adaptation from a Theoretical Perspective (CVPR 2024 · similarity 0.92)
- Generalized Source-Free Domain Adaptation (ICCV 2021 · similarity 0.92)
- Revisiting Source-Free Domain Adaptation: Insights into Representativeness, Generalization, and Variety (CVPR 2025 · similarity 0.92)
- Diffusion-Driven Progressive Target Manipulation for Source-Free Domain Adaptation (NeurIPS 2025 · similarity 0.90)
- SF(DA)$^2$: Source-free Domain Adaptation Through the Lens of Data Augmentation (ICLR 2024 · 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.
