Research map / Detection and segmentation / Open-set and few-shot detection
Reinforced active learning for image segmentation
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 |
|---|---|---|---|
| use the RL-based region selector instead of entropy sampling | active learning for semantic segmentation · Cityscapes validation, 12k regions | mIoU: 62.29 -> 63.32 | Helped |
| use the RL-based region selector instead of uniform sampling | active learning for semantic segmentation · Cityscapes validation, 12k regions | mIoU: 58.78 -> 63.32 | Helped |
| use the RL-based region selector instead of BALD | active learning for semantic segmentation · Cityscapes validation, 12k regions | mIoU: 60.64 -> 63.32 | Helped |
| add three pooled entropy maps rather than using only max-pooled entropy (ablation) | Ours state and action representation · Cityscapes validation, budget 4K | mIoU: 57.62 ± 0.96 -> 57.65 ± 1.30 | Helped |
| add KL-distance distributions comparing candidate regions with labeled and unlabeled regions (ablation) | Ours state and action representation · Cityscapes validation, budget 4K | mIoU: 57.65 ± 1.30 -> 59.18 ± 0.62 | Helped |
| label regions rather than full images (ablation) | active learning baselines and Ours · CamVid validation, budget 0.5k | mIoU: 69.44 ± 0.22 -> 71.85 ± 0.68 | Helped |
| select 24 regions per step instead of 1 (ablation) | Ours · CamVid validation, budget 0.5k | mIoU: 71.10 ± 0.75 -> 71.85 ± 0.68 | Helped |
| select 24 rather than 12, 36, 48, or 72 regions per step (ablation) | Ours · CamVid validation, budget 0.5k | mIoU: 70.93 ± 0.70; 71.24 ± 0.49; 71.25 ± 1.17; 71.20 ± 0.53 -> 71.85 ± 0.68 | Mixed |
| optimize with an average of Q-values over all sub-actions (ablation) | batch-mode DQN · active learning for semantic segmentation | segmentation performance | Hurt |
| increase the number of regions selected per step (ablation) | batch-mode DQN · active learning for semantic segmentation | segmentation performance | No clear effect |
| use BALD rather than uniform sampling | active learning for semantic segmentation · CamVid | segmentation performance | Hurt |
| use the RL-based selector rather than entropy sampling | active learning for semantic segmentation · Cityscapes | additional labeled regions to reach the same performance: roughly 30% less additional labeled data | Helped |
Closest papers by meaning
- Active Learning for Semantic Segmentation with Multi-class Label Query (NeurIPS 2023 · similarity 0.90)
- Diffusion-Driven Two-Stage Active Learning for Low-Budget Semantic Segmentation (NeurIPS 2025 · similarity 0.90)
- Towards Fewer Annotations: Active Learning via Region Impurity and Prediction Uncertainty for Domain Adaptive Semantic Segmentation (CVPR 2022 · similarity 0.89)
- Guideline-Consistent Segmentation via Multi-Agent Refinement (AAAI 2026 · similarity 0.88)
- Pixels to Tokens: Token Space Efficient Active Learning for Low-Budget Semantic Segmentation (NeurIPS 2026 · similarity 0.88)
- Active Label Correction for Semantic Segmentation with Foundation Models (ICML 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.