Research map / Label-efficient and robust learning / Domain adaptation and generalization
Density Matters: Improved Core-Set for Active Domain Adaptive Segmentation
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 a density-aware Core-set selection for active domain adaptive semantic segmentation, using a tighter Core-set bound based on average radial distance, a dynamic masked convolution density proxy, and a density-aware greedy sampler.
Problem
Active domain adaptation for semantic segmentation must choose few pixel labels under tight budgets; existing uncertainty/diversity methods ignore local feature-space density, and the classical Core-set bound is loose for dense pixel candidates.
Approach
Derives a tighter Core-set loss bound in terms of the expected average radial distance over each selected point's coverage area, estimates per-candidate coverage density with a Dynamic Masked Convolution proxy, and performs greedy selection that rescales distances by density.
Key results
- On GTAV to Cityscapes with DeepLab V2 at 2.2% budget, mIoU 71.1 vs RIPU 69.6 and D2ADA 69.7 (fully supervised 71.3).
- With DeepLab V3+ at 5% budget, mIoU 72.2 on GTAV and 73.2 on SYNTHIA (vs D2ADA 71.3 and 72.7).
- Ablation: density-aware greedy with DMC improves over K-Center Greedy by 1.7 mIoU on GTAV (71.1 vs 69.4 at 2.2%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Density-aware Core-set selection instead of RIPU | DeepLab V2 · GTAV → Cityscapes, 2.2% labels | mIOU: over 1.4 mIOU | Helped |
| Use Density-aware Core-set selection instead of D2ADA | DeepLab V2 · SYNTHIA → Cityscapes, 2.2% labels | mIOU: over 1.4 mIOU | Helped |
| Add DMC to Density-aware Greedy (ablation) | Density-aware Greedy · GTAV → Cityscapes, 2.2% labels | mIOU: 70.4 -> 71.1 | Helped |
| Add DMC to Density-aware Greedy (ablation) | Density-aware Greedy · SYNTHIA → Cityscapes, 2.2% labels | mIOU: 71.4 -> 72.1 | Helped |
| Use Density-aware Greedy instead of K-Center Greedy (ablation) | Core-set selection · GTAV → Cityscapes, 2.2% labels | mIOU: 69.4 -> 71.1 | Helped |
| Use Density-aware Greedy instead of K-Center Greedy (ablation) | Core-set selection · SYNTHIA → Cityscapes, 2.2% labels | mIOU: 70.7 -> 72.1 | Helped |
| Add DMC as feature regularization to K-Center Greedy (ablation) | K-Center Greedy · GTAV → Cityscapes, 2.2% labels | mIOU: 69.4 -> 70.1 | Helped |
| Add DMC as feature regularization to K-Center Greedy (ablation) | K-Center Greedy · SYNTHIA → Cityscapes, 2.2% labels | mIOU: 70.7 -> 71.1 | Helped |
| Replace 3×3 DMC with 5×5 DMC (ablation) | Density-aware Greedy · GTAV → Cityscapes, 2.2% labels | mIOU: 70.4 -> 71.1 | Helped |
| Replace 3×3 DMC with 5×5 DMC (ablation) | Density-aware Greedy · SYNTHIA → Cityscapes, 2.2% labels | mIOU: 71.5 -> 72.1 | Helped |
| Use Density-aware Greedy instead of K-Center Greedy | Core-set selection · Cityscapes training set | Core-set Loss: 0.646 -> 0.550 | Helped |
| Use our method instead of RAND | Active learning selection · GTAV → Cityscapes, 2.2% labels | mIOU: 7.3 | Helped |
Surprising findings
- The K-Center Greedy solution yields a smaller plain radius delta but a larger max average radial distance, so the radius-based bound is not the best proxy for the actual Core-set loss.
Limitations
- Relies on a candidate pre-filtering step using source-domain margin loss, which requires source data.
- Hyperparameters beta and tau are tuned on the training data to reduce the bound.
- Gains over strongest baselines at high budgets are modest (about 1 mIoU) and the method is evaluated only on two synthetic-to-real benchmarks.
Experimental setup
Datasets: GTAV SYNTHIA Cityscapes
Benchmarks: GTAV to Cityscapes SYNTHIA to Cityscapes
Metrics: mIoU
Baselines: Source Only AdaptSeg CBST CRST FDA TPLD IAST DPL-Dual BAPA-Net ProDA WeakDA LabOR
Builds on: Core-set (Sener & Savarese) k-Center Greedy RIPU learned image compression context model MIC
Closest papers by meaning
- Towards Fewer Annotations: Active Learning via Region Impurity and Prediction Uncertainty for Domain Adaptive Semantic Segmentation (CVPR 2022 · similarity 0.90)
- On the Road to Online Adaptation for Semantic Image Segmentation (CVPR 2022 · similarity 0.89)
- Diffusion-Driven Two-Stage Active Learning for Low-Budget Semantic Segmentation (NeurIPS 2025 · similarity 0.89)
- Open-Set Domain Adaptation for Semantic Segmentation (CVPR 2024 · similarity 0.88)
- Balanced Learning for Domain Adaptive Semantic Segmentation (ICML 2025 · similarity 0.87)
- Stable Neighbor Denoising for Source-free Domain Adaptive Segmentation (CVPR 2024 · 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.