Research map / Detection and segmentation / CNN features
Boosting the Dual-Stream Architecture in Ultra-High Resolution Segmentation with Resolution-Biased Uncertainty Estimation
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 resolution-biased uncertainty estimator (entropy map plus high-frequency residual) in the low-resolution stream that drives patch selection, uncertainty-weighted fusion, and loss weighting for dual-stream ultra-high-resolution segmentation.
Problem
Ultra-high-resolution semantic segmentation must balance accuracy against inference cost; existing dual-stream methods focus on one of three objectives (ensemble, selective zoom, complementary learning) in isolation and ignore resolution-induced prediction uncertainty.
Approach
Estimate resolution-biased uncertainty in the low-resolution stream by combining the entropy of the coarse prediction with a high-frequency feature residual (Laplacian-style), then reuse this single uncertainty map to select UHR patches, weight fusion between streams, and weight the UHR-stream loss.
Key results
- On DeepGlobe, the full model reaches 75.1 mIoU at 44.38 FPS and 1638 MB, versus WSDNet's 74.1 mIoU at 30.30 FPS.
- On Inria Aerial, the full model reaches 79.0 mIoU, about 2.5 points above GPWFormer (76.5) and 1.8 above SGHRQ (77.3)'s reported 77.3 is exceeded; the K=50% variant still gives 76.3 mIoU at 14.01 FPS.
- Uncertainty-aware patch selection at K=50% on Inria Aerial lowers mIoU loss to 76.3 from 79.0 versus 62.9 with random selection.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the uncertainty-aware ensembler (ablation) | UHR segmentation baseline · Inria Aerial | mIoU: 73.5 -> 76.1 | Helped |
| add the uncertainty-based complementer (ablation) | UHR segmentation with UAE · Inria Aerial | mIoU: 76.1 -> 78.5 | Helped |
| add binary-mask weighting (ablation) | UHR segmentation with UAE and UBC · Inria Aerial | mIoU: 78.5 -> 79.0 | Helped |
| use uncertainty-aware rather than random patch selection (ablation) | UHR segmentation, 50% selection · Inria Aerial | mIoU: 62.9 -> 76.3 | Helped |
| combine entropy and resolution-bias estimates rather than using entropy alone (ablation) | UHR segmentation · Inria Aerial | mIoU: 77.7 -> 79.0 | Helped |
| combine entropy and resolution-bias estimates rather than using resolution-bias estimation alone (ablation) | UHR segmentation · Inria Aerial | mIoU: 76.2 -> 79.0 | Helped |
| combine entropy and resolution-bias estimates rather than using entropy alone (ablation) | UHR segmentation, K=75% · Inria Aerial | mIoU: 76.5 -> 78.2 | Helped |
| combine entropy and resolution-bias estimates rather than using resolution-bias estimation alone (ablation) | UHR segmentation, K=75% · Inria Aerial | mIoU: 74.9 -> 78.2 | Helped |
| combine entropy and resolution-bias estimates rather than using entropy alone (ablation) | UHR segmentation, K=50% · Inria Aerial | mIoU: 74.8 -> 76.3 | Helped |
| combine entropy and resolution-bias estimates rather than using resolution-bias estimation alone (ablation) | UHR segmentation, K=50% · Inria Aerial | mIoU: 73.0 -> 76.3 | Helped |
| prune input patches to K=75% (ablation) | Ours · DeepGlobe test | mIoU: 75.1 -> 74.5 | Hurt |
| prune input patches to K=50% (ablation) | Ours · DeepGlobe test | mIoU: 75.1 -> 73.7 | Hurt |
Surprising findings
- Reported FPS speedups over MagNet (about 55x on DeepGlobe) come from avoiding iterative patch processing, and the uncertainty estimator adds very little cost.
- On URUR, the 'others' class is nearly absent, so mIoU gains are underestimated; the authors suggest excluding it from evaluation.
Limitations
- Binary patch pruning only; more granular hierarchical grouping not explored.
- Gains are smaller on DeepGlobe because errors there stem from category confusion rather than resolution loss.
- Uses fixed hyperparameters (alpha, th1, th2 = 0.5) chosen for simplicity without a broad sensitivity study.
Experimental setup
Datasets: DeepGlobe Inria Aerial URUR ISIC CRAG
Benchmarks: DeepGlobe Inria Aerial URUR ISIC CRAG
Metrics: mIoU accuracy FPS GPU memory
Baselines: ISDNet WSDNet SGHRQ GPWFormer FCtL GLNet PPN MagNet CascadePSP STDC
Builds on: GLNet ISDNet WSDNet SGHRQ PPN
Closest papers by meaning
- Ultra-High Resolution Segmentation via Boundary-Enhanced Patch-Merging Transformer (AAAI 2025 · similarity 0.86)
- Uncertainty-aware Fine-tuning of Segmentation Foundation Models (NeurIPS 2024 · similarity 0.85)
- Diffusion-Driven Two-Stage Active Learning for Low-Budget Semantic Segmentation (NeurIPS 2025 · similarity 0.85)
- ValUES: A Framework for Systematic Validation of Uncertainty Estimation in Semantic Segmentation (ICLR 2024 · similarity 0.85)
- UncertainSAM: Fast and Efficient Uncertainty Quantification of the Segment Anything Model (ICML 2025 · similarity 0.85)
- Toward Real Ultra Image Segmentation: Leveraging Surrounding Context to Cultivate General Segmentation Model (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.