Research map / Label-efficient and robust learning / Recognition and long-tail
RankED: Addressing Imbalance and Uncertainty in Edge Detection Using Ranking-based Losses
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 RankED, an edge detector trained with ranking-based losses (differentiable AP loss for positive-over-negative ranking plus a sorting loss ordering positives by annotator-agreement certainty) on a Swin-Transformer backbone.
Problem
Edge detection suffers from heavy positive/negative pixel imbalance and label uncertainty from annotator disagreement, which standard class-balanced cross-entropy, dice loss, and majority-agreement thresholding handle poorly.
Approach
Replaces score-based edge losses with ranking-based losses: an AP-approximating loss ranks positive pixels above negatives, and a sorting loss (adapted from RS Loss) ranks positives by a pixel-level label certainty map computed from multiple annotations, so high-certainty edges receive more priority.
Key results
- On NYUD-v2 RGB, RankED reports 78.0% ODS, 79.3% OIS and 82.6% AP, the best of the compared methods.
- On Multi-cue edge detection, RankED (R+S) reaches 96.2 ODS, 96.5 OIS and 97.3 AP, versus 89.5 ODS for the prior best reported.
- On BSDS500 (SS), RankED (R+S) reports AP 0.895 and MS AP 0.911, with the sorting term adding gains over ranking alone; vectorized ranking loss is about 100x faster than the official AP Loss implementation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add sorting loss for positive pixels by label certainty (ablation) | RANKED · BSDS500 test | AP: .886 -> .895 | Helped |
| add sorting loss for positive pixels by label certainty (ablation) | RANKED · Multi-cue test, edge | AP: .962±.004 -> .973±.006 | Helped |
| add sorting loss for positive pixels by label certainty (ablation) | RANKED · Multi-cue test, boundary | AP: .992 ±.002 -> .995±.001 | Helped |
| use ranking loss instead of class-balanced cross-entropy (ablation) | RANKED · NYUD-v2 test, RGB | AP: .802 -> .826 | Helped |
| use ranking loss instead of CE + Dice (ablation) | RANKED · BSDS500 test | AP: .872 -> .886 | Helped |
| use the proposed certainty map instead of pixel-wise averaged labels (ablation) | RANKED · BSDS500 test | AP: 0.880 -> 0.895 | Helped |
| weight pixel loss by (1 − ci) to prioritize uncertain pixels (ablation) | RANKED · BSDS500 test | AP: 0.895 -> 0.849 | Hurt |
| use ranking and sorting losses instead of the strongest listed baseline | RANKED · BSDS500 test, single-scale | AP: .892 -> .895 | Helped |
| use ranking and sorting losses instead of the strongest listed baseline | RANKED · BSDS500 test, multi-scale | AP: .897 -> .911 | Helped |
| use ranking loss instead of the strongest listed baseline | RANKED · BSDS500 test, single-scale + VOC | AP: .902 -> .901 | Hurt |
| use ranking loss instead of the strongest listed baseline | RANKED · NYUD-v2 test, RGB | AP: .797 -> .826 | Helped |
| use ranking and sorting losses instead of the strongest listed baseline | RANKED · Multi-cue test, edge | AP: .949±.002 -> .973±.006 | Helped |
Surprising findings
- Replacing the certainty map with an uncertainty-weighting scheme from UAED (giving more weight to uncertain pixels) substantially drops performance (ODS 0.789 vs 0.824 on BSDS).
- Only one exception to AP superiority on BSDS: at SS+VOC, RankED is 0.1 AP behind UAED.
Limitations
- Fully-vectorized implementation needs at least 45 GB GPU memory at 320x320; semi-vectorized needs 20 GB.
- Sorting loss cannot be applied to single-annotation datasets such as NYUD-v2 and Pascal Context, and the paper notes ODS/OIS are on par with others rather than best.
- Results depend on hand-tuned hyperparameters (delta, alpha, distance tolerance d, learning rate) per dataset.
Experimental setup
Datasets: BSDS500 NYUDv2 Multi-cue Pascal Context
Benchmarks: BSDS500 NYUDv2 Multi-cue
Metrics: ODS OIS AP uncertainty-aware ODS/OIS/AP
Baselines: HED BDCN PiDiNet EDTER UAED ACTD FCL-Net CHRNet LDC RCF
Builds on: AP Loss RS Loss class-balanced cross-entropy UperNet
Closest papers by meaning
- The Treasure Beneath Multiple Annotations: An Uncertainty-Aware Edge Detector (CVPR 2023 · similarity 0.89)
- SAUGE: Taming SAM for Uncertainty-Aligned Multi-Granularity Edge Detection (AAAI 2025 · similarity 0.87)
- Robust Visual Recognition with Class-Imbalanced Open-World Noisy Data (AAAI 2024 · similarity 0.84)
- DiffusionEdge: Diffusion Probabilistic Model for Crisp Edge Detection (AAAI 2024 · similarity 0.83)
- RankSEG-RMA: An Efficient Segmentation Algorithm via Reciprocal Moment Approximation (NeurIPS 2025 · similarity 0.83)
- Learning to Rank Patches for Unbiased Image Redundancy Reduction (CVPR 2024 · similarity 0.83)
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.