MatchED: Crisp Edge Detection Using End-to-End, Matching-based Supervision

CVPR 2026

Bedrettin Cetinkaya, Sinan Kalkan*, Emre Akbas*
Middle East Technical University and METU ROMER Robotics Center *(Equal senior authorship)
Paper arXiv Code

Overview

MatchED Overview
(a) Existing edge detectors require the use of a hand-crafted post-processing stage (e.g., NMS and thinning). (b) With MatchED, we generate a one-pixel-wide edge map in an end-to-end trainable manner. The recent DiffusionEdge’s results (c) improve when integrated with MatchED (d).

Our proposed framework, MatchED, is a lightweight, plug-and-play supervision module that adds only ~21K parameters. It can be seamlessly appended to any edge detection model, enabling joint, end-to-end learning of crisp edges.

Quantitative Results

MatchED significantly outperforms state-of-the-art crisp edge detection methods in terms of the Average Crispness (AC) metric on three benchmarks: BSDS, Multi-Cue, and BIPED.

Method BSDS Multi-Cue BIPED
PiDiNet + Dice Loss .306 .208 .340
PiDiNet + Tracing Loss .333 .217 .296
PiDiNet + Label Refinement .424 .424 .512
DiffusionEdge .476 .462 .849
PiDiNet + MatchED .930 .810 .941

Visual Results

MatchED delivers a substantial improvement in edge crispness across diverse architectures and datasets. When integrated as a plug-and-play module into models like DiffusionEdge or PiDiNet, it consistently produces sharper edges without using traditional post-processing.

BSDS DiffusionEdge Comparison
Qualitative comparisons on BSDS dataset using DiffusionEdge. We show results from DiffusionEdge (official checkpoint), raw outputs from our pipeline, raw outputs after applying NMS, and their corresponding MatchED integrated results, respectively. Best viewed zoomed-in.
NYUD PiDiNet Comparison
Qualitative comparisons on NYUD dataset using PiDiNet. We show results from PiDiNet (official checkpoint), raw outputs from our pipeline, raw outputs after applying NMS, and their corresponding MatchED integrated results, respectively. Best viewed zoomed-in.

Acknowledgements

This work was supported by the Council of Higher Education Research Universities Support program through METU Scientific Research Projects (``New Techniques in Visual Recognition’’, Project No. ADEP-312-2024-11485). We also gratefully acknowledge the computational resources provided by METU-ROMER, Center for Robotics and Artificial Intelligence, as well as TUBITAK ULAKBIM Truba.

BibTeX

@article{cetinkaya2026matched,
  author    = {Bedrettin Cetinkaya and Sinan Kalkan and Emre Akbas},
  title     = {MatchED: Crisp Edge Detection Using End-to-End, Matching-based Supervision},
  year={2026},
  booktitle={IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  url={https://cvpr26-matched.github.io/}
}