Research map / Label-efficient and robust learning / Domain adaptation and generalization
Robust Nearest Neighbors for Source-Free Domain Adaptation under Class Distribution Shift
Antonio Tejero-de-Pablos, Riku Togashi, Mayu Otani, Shin'ichi Satoh
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 robust nearest neighbors with a Swin-B generic model | PL guided · VisDA-C RSUT (SFDA) | class-wise mean accuracy: 83.59 -> 88.84 | Improved |
| add robust nearest neighbors with a ResNet generic model on Office-Home | PL guided · Office-Home RSUT (SFDA) | class-wise mean accuracy: 61.05 -> 59.67 | Hurt |
| add robust nearest neighbors with a ResNet50 generic model on DomainNet | PL guided · DomainNet subset (SFDA) | class-wise mean accuracy: 80.12 -> 72.74 | Hurt |
| add robust nearest neighbors with ViT-B on DomainNet | PL guided · DomainNet subset (SFDA) | class-wise mean accuracy: 80.12 -> 83.9 | Improved |
| use Swin-B generic model instead of ViT-B on DomainNet (ablation) | PL guided + Ours · DomainNet subset (SFDA) | class-wise mean accuracy: 83.9 -> 81.4 | Hurt |
| add robust nearest neighbors with Swin-B to basic pseudolabeling | PL base · VisDA-C RSUT (SFDA) | class-wise mean accuracy: 81.01 -> 86.64 | Improved |
| set common-neighbor range K to 100 rather than 10 or 500 (ablation) | Robust NN · VisDA-C RSUT | class-wise mean accuracy: 83.96 (K=10), 83.56 (K=500) -> 86.6 (K=100) | Improved |
| add robust nearest neighbors to test-time adaptation pseudolabeling with Swin-B | Pseudolabel TTA · Office-Home RSUT (TTA) | class-wise mean accuracy: 52.34 -> 60.16 | Improved |
| add robust nearest neighbors to TTA on VisDA-C versus label shift adapter | Pseudolabel TTA · VisDA-C RSUT (TTA) | class-wise mean accuracy: 47.12 -> 52.49 (vs 72.97 for label shift adapter) | Mixed |
| increase class distribution shift severity (ablation) | PL guided + Ours · VisDA-C RSUT | class-wise mean accuracy: ViT-B: 86.72 -> 83.88; ResNet101: 86.60 -> 80.46 | Hurt |
Closest papers by meaning
- Source-Free Domain Adaptation via Distribution Estimation (CVPR 2022 · similarity 0.91)
- Understanding and Improving Source-free Domain Adaptation from a Theoretical Perspective (CVPR 2024 · similarity 0.91)
- BMD: A General Class-Balanced Multicentric Dynamic Prototype Strategy for Source-Free Domain Adaptation (ECCV 2022 · similarity 0.90)
- Source-Free Domain Adaptation Guided by Vision and Vision-Language Pre-training (IJCV 2024 · similarity 0.90)
- Revisiting Source-Free Domain Adaptation: Insights into Representativeness, Generalization, and Variety (CVPR 2025 · similarity 0.89)
- SF(DA)$^2$: Source-free Domain Adaptation Through the Lens of Data Augmentation (ICLR 2024 · 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.
