Research map / Label-efficient and robust learning / Domain adaptation and generalization
Hyperbolic Active Learning for Semantic Segmentation under Domain Shift
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 HALO, a hyperbolic-network active domain adaptation method for semantic segmentation that scores pixels by the product of hyperbolic radius and prediction entropy as a proxy for epistemic uncertainty.
Problem
Pixel-level active domain adaptation for semantic segmentation under domain shift, where existing acquisition strategies rely mainly on prediction entropy and miss data scarcity as a source of epistemic uncertainty.
Approach
Interprets the hyperbolic radius of pixel embeddings (hierarchy-free hyperbolic MLR segmenter) as an indicator of data scarcity, and combines it with prediction entropy into an acquisition score approximating epistemic uncertainty; also adds Hyperbolic Feature Reweighting to stabilize hyperbolic training.
Key results
- HALO reaches 74.5% mIoU on GTAV→Cityscapes with 5% budget, surpassing the supervised DA baseline of 71.9% and prior RIPU at 71.2%.
- HALO achieves 75.6% mIoU on SYNTHIA→Cityscapes with 5% budget (+4.2% over RIPU per the paper) and 63.2% on Cityscapes→ACDC is reported for 2.2% budget vs RIPU 62.3%.
- Ablation on GTAV→CS with 5% budget: entropy alone 63.2%, radius alone 64.1%, product 74.5%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HALO instead of RIPU with a 2.2% label budget | GTAV→Cityscapes · GTAV→Cityscapes, 2.2% budget | mIoU: +1.2% mIoU | Helped |
| Use HALO instead of the strongest reported active-learning baseline with a 5% budget | GTAV→Cityscapes · GTAV→Cityscapes, 5% budget | mIoU: +3.3% mIoU | Helped |
| Use HALO with 5% labels instead of supervised domain adaptation with 100% labels | GTAV→Cityscapes · GTAV→Cityscapes | mIoU: +2.6% | Helped |
| Use HALO instead of the strongest reported active-learning baseline | SYNTHIA→Cityscapes · SYNTHIA→Cityscapes, 2.2% and 5% budgets | mIoU: +2.4% and +4.2% | Helped |
| Use HALO instead of RIPU with a 5% label budget | Cityscapes→ACDC · Cityscapes→ACDC, 5% budget | mIoU: +2.9% mIoU | Helped |
| Combine hyperbolic radius and prediction entropy rather than using either alone (ablation) | HALO · GTAV→Cityscapes, 5% budget | mIoU: 63.2% and 64.1% -> 74.5% | Helped |
| Add HFR to the hyperbolic model for source+target adaptation (ablation) | HALO · GTAV→Cityscapes, source+target | mIoU: 72.9 -> 74.5 | Helped |
| Add HFR to the hyperbolic model in source-only training (ablation) | Hyperbolic DeepLab-v3+ · GTAV→Cityscapes, source-only | mIoU: 39.0 -> 38.9 | No clear effect |
| Add HFR to the Euclidean model (ablation) | DeepLab-v3+ · GTAV→Cityscapes, source-only | mIoU: 36.3 -> 22.7 | Hurt |
| Use region-based rather than pixel-based acquisition (ablation) | HALO · GTAV→Cityscapes, 5% budget | mIoU: 74.5 -> 74.1 | Hurt |
| Replace HFR with Feature Clipping (ablation) | Hyperbolic HALO · GTAV→Cityscapes, 5% budget | mIoU: 74.5 -> 73.3 | Hurt |
| Select only ground-truth boundary pixels instead of using RIPU's pseudo-label boundary selection (ablation) | RIPU · GTAV→Cityscapes | mIoU: 71.2 -> 69.8 | Hurt |
Surprising findings
- Hyperbolic radius correlates with class pixel frequency (rho=-0.899) rather than with accuracy, contrary to the uncertainty interpretation in prior hyperbolic metric learning work.
- Boundary-only selection from ground-truth boundaries performs worse than RIPU's pseudo-label boundary selection in an oracle experiment (69.8 vs 71.2).
Limitations
- Lacks rigorous mathematical validation of the hyperbolic radius properties.
- Relies on a source model pretrained on synthetic data, requiring synthetic-to-real adaptation.
- Pixel-level manual annotation remains time-consuming.
Experimental setup
Datasets: GTAV SYNTHIA Cityscapes ACDC
Benchmarks: GTAV to Cityscapes SYNTHIA to Cityscapes Cityscapes to ACDC
Metrics: mIoU
Baselines: RIPU LabOR AADA MADA D2ADA Euclidean Supervised DA Hyperbolic Supervised DA
Builds on: RIPU Hyperbolic Neural Networks DeepLab-v3+ Hyperbolic multinomial logistic regression
Closest papers by meaning
- Diffusion-Driven Two-Stage Active Learning for Low-Budget Semantic Segmentation (NeurIPS 2025 · similarity 0.88)
- Pixels to Tokens: Token Space Efficient Active Learning for Low-Budget Semantic Segmentation (NeurIPS 2026 · similarity 0.86)
- Density Matters: Improved Core-Set for Active Domain Adaptive Segmentation (AAAI 2024 · similarity 0.85)
- Hyperbolic Uncertainty-Aware Few-Shot Incremental Point Cloud Segmentation (CVPR 2025 · similarity 0.85)
- Understanding Fine-tuning CLIP for Open-vocabulary Semantic Segmentation in Hyperbolic Space (CVPR 2025 · similarity 0.84)
- Hyperbolic Embeddings of Supervised Models (NeurIPS 2024 · 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.