Research map / Detection and segmentation / CNN features
MuGE: Multiple Granularity Edge Detection
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 MuGE, an edge detector that predicts multiple edge maps at controllable granularities by encoding annotation granularity with a learned classifier and injecting it into spatial and high-frequency features of a UAED-based encoder-decoder.
Problem
Edge detection is subjective and ambiguous, but most deterministic detectors output only one edge map per image, and probabilistic methods give limited, uncontrollable diversity.
Approach
Train a VGG16 granularity classifier on the simplest and most complex annotations to produce a scalar granularity in [0,1], then multiply it into half of the multi-scale features (spatial) and into the high-frequency DFT components, so varying the scalar at inference yields controllable edge maps.
Key results
- On BSDS500 MS-VOC setting, MuGE (M=11) reaches ODS 0.861, OIS 0.867, AP 0.909, surpassing UAED (ODS 0.844).
- On Multicue boundary, MuGE achieves ODS 0.875, OIS 0.879, AP 0.932, a new state of the art over UAED (0.864 ODS).
- LPIPS diversity on BSDS500 rises from 0.0118 for UAED to 0.1065 for MuGE at M=11.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Generate multiple predictions and select the best-matching map | MuGE · BSDS500 MS-VOC | ODS: 0.844 -> 0.861 | Helped |
| Generate multiple predictions and select the best-matching map | MuGE · Multicue boundary | ODS: 0.864 -> 0.875 | Helped |
| Add edge-granularity embedding in the frequency domain (ablation) | UAED-based detector · BSDS500 | ODS: 0.829 -> 0.846 | Helped |
| Add spatial-domain edge-granularity embedding to frequency-domain embedding (ablation) | UAED-based detector · BSDS500 | ODS: 0.846 -> 0.846 | No clear effect |
| Add CLIP loss (ablation) | MuGE · BSDS500 | ODS: +0.3% | Helped |
| Add focal frequency loss (ablation) | MuGE · BSDS500 | Average Crispness: from 0.228 to 0.291 | Helped |
| Replace edge-granularity classifier with normalization (ablation) | MuGE · BSDS500 | ODS: 0.850 -> 0.847 | Hurt |
| Concatenate edge granularity with the input image instead of feature embedding (ablation) | MuGE · BSDS500 | ODS: 0.850 -> 0.849 | Hurt |
| Replace feature embedding with AdaIN (ablation) | MuGE · BSDS500 | ODS: 0.850 -> 0.804 | Hurt |
| Vary the edge-granularity interval from 0.5 to 0.1, increasing predictions from M = 3 to M = 11 | MuGE · BSDS500 single-scale | ODS: 0.845 -> 0.850 | Helped |
| Increase generated edge maps from M = 3 to M = 11 | MuGE · BSDS500 | LPIPS: 0.1663; 0.1065 | Mixed |
Surprising findings
- Using a fixed simple granularity (alpha=0) yields lower scores than middle values, since ground truth merges all annotations.
Limitations
- Granularity is controlled only through annotation-derived encoding; other control schemes such as text prompts are not explored.
- Single-scale gains over UAED on ODS are modest for some settings (e.g., FFL loss alone gives little improvement).
Experimental setup
Datasets: BSDS500 PASCAL VOC Context Multicue
Benchmarks: BSDS500 Multicue
Metrics: ODS OIS AP LPIPS average crispness
Baselines: UAED EDTER BDCN RCF LDC BetaNet DSCD Canny Multicue VR
Builds on: UAED FFL (focal frequency loss) EfficientNet UNet++ CLIP
Closest papers by meaning
- SAUGE: Taming SAM for Uncertainty-Aligned Multi-Granularity Edge Detection (AAAI 2025 · similarity 0.85)
- GraCo: Granularity-Controllable Interactive Segmentation (CVPR 2024 · similarity 0.83)
- DiffusionEdge: Diffusion Probabilistic Model for Crisp Edge Detection (AAAI 2024 · similarity 0.82)
- Multi-Granularity Video Object Segmentation (AAAI 2025 · similarity 0.82)
- ScaleLSD: Scalable Deep Line Segment Detection Streamlined (CVPR 2025 · similarity 0.81)
- RankED: Addressing Imbalance and Uncertainty in Edge Detection Using Ranking-based Losses (CVPR 2024 · similarity 0.81)
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.