Research map / Label-efficient and robust learning / Domain adaptation and generalization
Discriminative Pattern Calibration Mechanism for Source-Free Domain Adaptation
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 Discriminative Pattern Calibration (DPC), which uses Grad-CAM attention calibrated by an autoencoder plus attention-reversed mixup to plug into existing SFDA methods (SHOT, AaD) and improve target accuracy.
Problem
Source-free domain adaptation suffers because distribution shift makes the source model rely on classification-irrelevant regions, producing invalid discriminative semantics on target images.
Approach
Use Grad-CAM attention maps, calibrated by an encoder-decoder conditioned on prediction confidence, to emphasize discriminative regions via semantic filling in feature space, and mix target images by removing the attended object region from one image (attention-reversed mixup) to improve classifier robustness.
Key results
- On Office-Home, SHOT+DPC reaches 73.2% average accuracy (+1.4 over SHOT's 71.8%), and AaD+DPC reaches 73.1% (+1.3 over AaD).
- On VisDA-C, AaD+DPC reaches 88.8% average accuracy, exceeding FixBi by 1.6%.
- On Office-31, SHOT+DPC reaches 89.8% (+1.2) and AaD+DPC 90.5% (+0.8) average accuracy.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add DPC to SHOT | SHOT · Office-Home (12-task average) | accuracy: 71.8 -> 73.2 | Helped |
| Add DPC to AaD | AaD · Office-Home (12-task average) | accuracy: 71.8 -> 73.1 | Helped |
| Add DPC to SHOT | SHOT · Office-31 (6-task average) | accuracy: 88.6 -> 89.8 | Helped |
| Add DPC to AaD | AaD · Office-31 (6-task average) | accuracy: 89.7 -> 90.5 | Helped |
| Add DPC to SHOT | SHOT · VisDA-C | accuracy: 82.9 -> 84.3 | Helped |
| Add DPC to AaD | AaD · VisDA-C | accuracy: 87.7 -> 88.8 | Helped |
| Add the DSE module (ablation) | AaD · Office-Home: Cl→Ar, Pr→Ar, and Rw→Ar | accuracy | Helped |
| Add attention-induced mixture (AIM), using Grad-CAM attention without attention adjustment or semantic filling (ablation) | AaD · Office-Home: Cl→Ar, Pr→Ar, and Rw→Ar | accuracy | Helped |
| Replace attention-guided mixture with simple blending of two target images (ablation) | AaD · Office-Home: Cl→Ar, Pr→Ar, and Rw→Ar | accuracy | Hurt |
Surprising findings
- Ablation shows the DSE calibration module contributes more gain than the attention-induced mixture module.
Limitations
- Gains over SHOT and AaD are modest (about 0.8 to 1.4 points average).
- Relies on Grad-CAM attention maps whose quality depends on the source model's predictions and pseudo labels.
Experimental setup
Datasets: Office-31 Office-Home VisDA-C
Benchmarks: Office-31 Office-Home VisDA-C
Metrics: classification accuracy
Baselines: SHOT AaD NRC A2Net SFDA FixBi CDAN SRDC
Builds on: SHOT AaD Grad-CAM mixup
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 0.84)
- SF(DA)$^2$: Source-free Domain Adaptation Through the Lens of Data Augmentation (ICLR 2024 · similarity 0.84)
- Invertible Projection and Conditional Alignment for Multi-Source Blended-Target Domain Adaptation (AAAI 2025 · similarity 0.84)
- Diffusion-Driven Progressive Target Manipulation for Source-Free Domain Adaptation (NeurIPS 2025 · similarity 0.84)
- Proxy Denoising for Source-Free Domain Adaptation (ICLR 2025 · similarity 0.84)
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.