SuperEdge: A Measurement-Systems Framework for Self-Supervised Edge Detection

SuperEdge is a self-supervised edge detector packaged with a measurement-systems evaluation framework. Beyond conventional ODS / OIS accuracy, this repository quantifies the four first-class indicators an instrumentation engineer applies to any measurand:

Indicator	What it measures	Where it lives
Accuracy	ODS, OIS, AP across BIPED, BIPEDv2, BSDS500, NYUD, BSDS-RIND	evaluate.py
Repeatability	$\sigma_r$ under homographic perturbation, with bootstrap-1000 95% CIs	scripts/repeatability_sweep.py
Robustness	F-measure under $\gamma \in [0.4, 2.0]$ illumination drift and $\sigma^2 \in [0, 0.05]$ additive Gaussian noise	scripts/robustness_sweep.py
Deployment footprint	Latency p50/p95, GPU memory, parameters, FLOPs, Pareto-domination across the foundation-model comparison set	scripts/bench_pareto.py

A GUM-aligned (ISO/IEC Guide 98-3, JCGM 100:2008) Type-A / Type-B uncertainty budget combines the dominant noise sources into a combined relative standard uncertainty of 6.9% for downstream traceability—see scripts/uncertainty_budget.py.

A 6-DoF arm grasp setup is included as a qualitative application demonstration of edge-guided bounding-box refinement on transparent and reflective objects. A fully-powered, paired-trial statistical evaluation of grasp success is reserved for follow-up work; the scripts/mcnemar_grasp.py skeleton is retained for that future study and accepts any paired-trial CSV that matches the schema in its docstring.

Reproducibility note. Every numerical claim in §§II-G, III-C, III-D, III-E, III-F of the accompanying T-IM submission is backed by exactly one script in scripts/, each with pinned seeds, a documented invocation block, and CSV/JSON output that matches a paper table or figure. See docs/REPRODUCE.md for the step-by-step pipeline from git clone to each numerical artifact.

Quick Start

git clone https://github.com/s20sc/SuperEdge.git
cd SuperEdge

# Training stack (matches the published checkpoint)
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

# Evaluation / measurement-systems stack (additional deps)
pip install -r requirements-eval.txt

# Reproduce paper Tables / Figures (see docs/REPRODUCE.md for full walkthrough)
python scripts/repeatability_sweep.py  --checkpoint export/<ckpt>.pth ...
python scripts/robustness_sweep.py     --checkpoint export/<ckpt>.pth ...
python scripts/bench_pareto.py         --model superedge ...
python scripts/uncertainty_budget.py   --repeatability results/repeatability.csv ...
python scripts/mcnemar_grasp.py        --trials data/grasp_trials/trials.csv ...

Repository Layout

.
├── model/                  # SuperEdge architecture (encoder + dual decoder)
├── dataset/                # COCO, BIPED, BSDS500 loaders + synthetic shapes
├── solver/                 # loss, NMS, BSDS-style F-measure evaluator
├── config/                 # Hydra-style YAML configs for each training stage
├── train.py                # 3-stage training pipeline
├── evaluate.py             # accuracy evaluation (ODS / OIS / AP)
├── homography_adaptation.py
├── object_level_label.py
├── scripts/                # T-IM measurement-systems evaluation (this release)
│   ├── repeatability_sweep.py   # §III-D, Table VII
│   ├── robustness_sweep.py      # §III-E, Tables VIII / IX
│   ├── uncertainty_budget.py    # §II-G, Table II
│   ├── bench_pareto.py          # §III-C, Table VI, Fig. 4
│   └── mcnemar_grasp.py         # §III-F (6-DoF arm)
├── data/
│   ├── manifests/
│   │   └── coco_repeatability_12.txt   # pinned 12-image scene list
│   └── grasp_trials/
│       └── trials.csv                  # released n=50 paired-trial log
├── docs/
│   └── REPRODUCE.md                    # step-by-step recipe per paper claim
├── requirements.txt              # training stack
├── requirements-eval.txt         # additional eval-only deps (scipy, statsmodels, fvcore, ...)
├── CITATION.cff
└── LICENSE                       # Apache-2.0

Pre-trained Weights

Download the BIPEDv2-trained checkpoint (1.29 M parameters):

# From the GitHub release page:
wget https://github.com/s20sc/SuperEdge/releases/download/v1.0/superedge_bipedv2.pth \
     -O export/superedge_bipedv2.pth

Citing

If you use this code or the released measurement protocol, please cite the T-IM paper:

@article{qin2026superedge,
  author  = {Qin, Xue and Leng, Kai and Zhang, Yuqi and Liu, Xin and
             Li, Tao and Chao, Pingfu and Li, Zhijun},
  title   = {{SuperEdge}: A Self-Calibrating Edge Detector
             for Vision-Based Measurement},
  journal = {IEEE Transactions on Instrumentation and Measurement},
  year    = {2026},
  note    = {In submission}
}

A machine-readable CITATION.cff is included.

License

Released under the Apache License 2.0. The trained checkpoints are released under the same terms.

Contact

Issues and PRs welcome on this repository. For research correspondence: Pingfu Chao (pfchao@suda.edu.cn) and Zhijun Li (lizhijun_os@hit.edu.cn), corresponding authors.