Research map / Label-efficient and robust learning / Domain adaptation and generalization
Dirichlet-based Uncertainty Calibration for Active Domain Adaptation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 Dirichlet-based evidential model with two-round uncertainty-origin selection (DUC) vs. prior active DA methods | active DA on miniDomainNet · miniDomainNet 5% budget | Accuracy (%) avg: +2.1 over EADA (72.9 vs 70.8) | Helped |
| Use DUC two-round selection (U_dis then U_data) instead of EADA | active DA on Office-Home · Office-Home 5% budget | Accuracy (%) avg: +1.3 over EADA (78.0 vs 76.7) | Helped |
| Replace random/entropy selection with U_dis-then-U_data selection in EDL (ablation) | EDL on Office-Home · Office-Home 5% budget | Accuracy (%) avg: EDL entropy 73.3 and random 71.1 vs variant A 74.1 | Helped |
| Add L_Udis loss (minimizing distribution uncertainty on unlabeled target data) (ablation) | variant A on Office-Home · Office-Home 5% budget | Accuracy (%) avg: 76.6 vs 74.1 | Helped |
| Add L_Udata loss on top of variant A (ablation) | variant A on Office-Home · Office-Home 5% budget | Accuracy (%) avg: 74.2 vs 74.1 | No clear effect |
| Add both L_Udis and L_Udata losses to EDL (variant D, no selection criterion) (ablation) | EDL on Office-Home · Office-Home 5% budget | Accuracy (%) avg: 68.6 vs 61.5 for EDL without selection (7.1 improvement) | Helped |
| Use U_dis in first round and U_data in second round (ordering) vs reversed (ablation) | DUC selection ordering · Office-Home and GTAV/SYNTHIA | Accuracy / mIoU: Office-Home 78.0 vs 77.5; GTAV 67.0 vs 66.2 | Helped |
| Replace DUC's U_dis/U_data with entropy-based criterion (Variant F with both losses) (ablation) | DUC on Office-Home · Office-Home 5% budget | Accuracy (%) avg: 76.7 vs 78.0 | Hurt |
| Use only U_dis (variant G) or only U_data (variant H) instead of the two-round selection (ablation) | DUC on Office-Home · Office-Home 5% budget | Accuracy (%) avg: 77.1 and 76.9 vs 78.0 | Hurt |
| Apply FixMatch-style pseudo-label loss on unlabeled target data (ablation) | DUC on Office-Home · Office-Home 5% budget | Accuracy (%) avg: 78.9 vs 78.0 | Helped |
| Replace evidential (Dirichlet) predictions with deterministic softmax in the uncertainty estimate | active DA · Office-Home ECE on target data | Expected calibration error: much lower ECE for DUC (value not given numerically) | Helped |
Closest papers by meaning
- Diffusion-Based Probabilistic Uncertainty Estimation for Active Domain Adaptation (NeurIPS 2023 · similarity 0.92)
- Revisiting the Domain Shift and Sample Uncertainty in Multi-source Active Domain Transfer (CVPR 2024 · similarity 0.91)
- Active Domain Adaptation with False Negative Prediction for Object Detection (CVPR 2024 · similarity 0.88)
- Pseudo-Calibration: Improving Predictive Uncertainty Estimation in Unsupervised Domain Adaptation (ICML 2024 · similarity 0.88)
- Category-Aware Active Domain Adaptation (ICML 2024 · similarity 0.87)
- Invertible Projection and Conditional Alignment for Multi-Source Blended-Target Domain Adaptation (AAAI 2025 · similarity 0.86)
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.