Pith. sign in

REVIEW 4 major objections 4 minor 38 references

Edge-Boundary-Texture Loss: A Tri-Class Generalization of Weighted Binary Cross-Entropy for Enhanced Edge Detection

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that replacing the standard weighted binary cross-entropy loss with a three-class loss—edge, boundary, and texture—makes deep edge detectors produce sharper edge maps, with weighted cross-entropy recovered as a limit case.

desk verdict A clean but shallow loss formulation whose strict-protocol gains are contradicted by the paper's own NMS results, so the abstract overclaims. read the letter →

arxiv 2507.06569 v1 pith:ECUFPTAI submitted 2025-07-09 cs.CV

Hao Shu This is my paper · ORCID
classification cs.CV
keywords edgedetectionlossfunctionweightedbinarycross-entropyedge-boundary-texturepixelclassificationboundarydeeplearningODS/OIS/AP
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that a single change to the training loss can make deep edge detectors produce sharper, more precise edges. It proposes the Edge-Boundary-Texture (EBT) loss, which sorts every pixel into one of three classes—true edge, boundary (close to an edge), or texture (distant background)—and gives each class its own weight in the cross-entropy objective. The author argues this is a principled generalization of the widely used Weighted Binary Cross-Entropy (WBCE) loss, since WBCE reappears when the boundary zone expands to cover the whole image. If correct, the claim matters because EBT is a drop-in replacement that reportedly improves average precision by double digits under a strict 1-pixel error tolerance without non-maximum suppression, across four models and five datasets with one fixed hyperparameter setting.

What carries the argument

The load-bearing mechanism is the three-class pixel taxonomy plus its weighting scheme. Edge pixels $\mathbb{Y}_E$ are the ground-truth edge pixels; boundary pixels $\mathbb{Y}_B$ are non-edge pixels inside a $(2r+1)\times(2r+1)$ square window centered on any edge pixel, with default radius $r=7$; texture pixels $\mathbb{Y}_T$ are everything else. The loss is a weighted binary cross-entropy sum in which each class receives a balancing coefficient ($B_E=1.0$, $B_B=0.8$, $B_T=0.5$ by default) times an adaptive size-balancing weight $W_E=|\mathbb{Y}_B\cup\mathbb{Y}_T|/|\mathbb{Y}|$, $W_B=|\mathbb{Y}_E\cup\mathbb{Y}_T|/|\mathbb{Y}|$, $W_T=|\mathbb{Y}_E\cup\mathbb{Y}_B|/|\mathbb{Y}|$. This targets extra supervision at the ambiguous band around edges while down-weighting distant texture, which is what produces thinner predictions.

What would settle it

Rerun the same four models on NYUD2 and BIPED2 under the traditional benchmark (resolution-scaled tolerance, e.g., 6–11 pixels, plus NMS) and compare AP and ODS between WBCE- and EBT-trained versions; if EBT does not beat WBCE there, the claimed superiority is confined to the strict no-NMS protocol. A second check is to train with boundary radii $r=3$ and $r=15$ in addition to $r=7$ on BSDS500 and measure whether AP gains are stable, which would test whether the fixed 7-pixel band is a meaningful semantic scale.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the blurriness of learned edge maps is not primarily an architecture problem but a supervision problem: WBCE treats every non-edge pixel identically, so the model receives no signal about the near-edge zone where misclassification actually happens. EBT corrects this by defining three pixel classes and weighting their cross-entropy terms separately, so the network learns to suppress responses just outside true edges instead of smearing them. The paper further proves that EBT contains WBCE as a limit case: as the boundary radius $r$ grows until the boundary class absorbs all non-edge pixels, and with edge and texture coefficients set appropriately, the EBT expression reduces exactly to the WBCE expression. Empirically, under the strictest evaluation (1-pixel tolerance, no NMS), EBT raises average AP from 0.418 to 0.484 on BIPED2 and from 0.226 to 0.299 on BSDS500, while keeping ODS and OIS gains around 1–3 percent. The appendix, however, reports the relaxed tolerance-plus-NMS benchmarks, where the advantage largely vanishes and EBT sometimes falls below WBCE, most consistently on NYUD2.

Load-bearing premise

The central experimental claim stands or falls on accepting the strict 1-pixel error tolerance without non-maximum suppression as the right evaluation protocol, because the paper's own appendix shows that under the standard relaxed tolerance with NMS, EBT frequently loses to WBCE, including all four AP comparisons on NYUD2.

Editorial extensions

If this is right

  • Because EBT is a drop-in replacement for WBCE, any existing CNN edge detector can switch objectives without architectural changes, keeping the same training schedule.
  • The average AP gains under 1-pixel no-NMS scoring are much larger than the ODS/OIS gains, implying EBT mainly suppresses false positives near edges rather than shifting overall detection quality.
  • The unified hyperparameter set ($B_E=1.0$, $B_B=0.8$, $B_T=0.5$) working across models and datasets suggests EBT is usable without per-dataset tuning.
  • The theoretical reduction to WBCE means EBT inherits WBCE's stability properties in the limit and should not introduce pathological training behavior when the boundary class is large.
  • Under the traditional relaxed tolerance with NMS, EBT's gains are inconsistent, so the practical benefit may be limited to settings where raw, un-thinned edge maps are wanted.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the mechanism is really about the near-edge band, then the optimal boundary radius should scale with image resolution; a testable extension is to set $r$ proportional to image size instead of the fixed value 7.
  • The large AP gains under strict scoring suggest EBT outputs need less non-maximum suppression; one could verify by counting local-maxima thickness in EBT versus WBCE predictions on the same test images.
  • The same tri-class reweighting idea transfers naturally to other thin-structure tasks such as lane marking, road boundary, or crack detection, where background pixels adjacent to the target are the main source of false positives.
  • The appendix's NYUD2 AP drops under the NMS protocol raise the possibility that EBT's benefit is concentrated in high-resolution, well-labeled edge datasets and may shrink when labels are derived from segmentation, as NYUD2's Canny-extracted labels are.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes the Edge-Boundary-Texture (EBT) loss, a tri-class reweighting of WBCE in which pixels are labeled as edge, boundary, or texture, with boundary pixels defined as non-edge pixels within a (2r+1)x(2r+1) window centered on edge pixels. The authors prove that WBCE arises as a limit case of EBT when the boundary region covers all non-edge pixels, and they report experiments on HED, BDCN, Dexi, and EES3 across five datasets under a strict 1-pixel error tolerance without NMS. The main-text tables show consistent AP gains under that protocol, and an ablation study on EES3/BRIND indicates moderate robustness to the boundary and texture weights. The appendix additionally reports results under traditional relaxed-tolerance-with-NMS benchmarks, where the picture changes substantially.

Significance. The idea of giving separate supervisory weight to a boundary band around edges is a plausible and clean extension of WBCE, and the asymptotic reduction to WBCE is correctly derived. The paper also tests four architectures and five datasets with a single hyperparameter setting, which is useful breadth for a loss-function study. However, the central empirical evidence is obtained under a nonstandard evaluation protocol that is sourced to the author's own unpublished preprint, and the paper's own appendix shows that EBT frequently loses under the standard NMS-based benchmarks. The unqualified claim of 'superiority ... both quantitatively and perceptually' is therefore not supported by the complete evidence in the manuscript. As submitted, the contribution is at most a narrowly scoped empirical finding about the strict no-NMS protocol, not a demonstrated general improvement.

major comments (4)
  1. [Abstract; Appendix Tables VI-X] The abstract's unqualified claim that EBT is superior 'both quantitatively and perceptually' is contradicted by the manuscript's own appendix. Under the traditional relaxed-tolerance-with-NMS protocol, Table X (NYUD2, 6-pixel tolerance) shows EBT reducing AP for every model: HED 0.648 vs 0.694, BDCN 0.683 vs 0.712, Dexi 0.687 vs 0.728, and EES3 0.719 vs 0.730. Table VI (BIPED2) shows Dexi-EBT AP 0.904 vs 0.923 and EES3-EBT AP 0.929 vs 0.932, and Table IX (UDED) shows HED-EBT and Dexi-EBT losing on ODS, OIS, and AP. The conclusion in Section V that 'EBT improves predictions, both quantitatively and qualitatively' is therefore not supported by the complete evidence; the claims must be scoped to the strict 1-pixel no-NMS protocol, or the appendix results must be reconciled.
  2. [Section IV.A.1; Section IV.B; Tables IV-V] The main-text evaluation is conducted exclusively under a 1-pixel error tolerance without NMS, a protocol attributed to the author's own unpublished preprint [19]. This is not a standard benchmark, and the paper provides no independent justification for adopting it as the primary protocol. Moreover, even under this preferred protocol, the gains are not uniform: Table IV reports HED-EBT on UDED with ODS 0.713 vs 0.716 and OIS 0.739 vs 0.751, and Table V reports BDCN-EBT on NYUD2 with ODS 0.402 vs 0.406, Dexi-EBT with ODS 0.405 vs 0.408, and an average NYUD2 ODS of 0.399 vs 0.401. The abstract's 'extensive experiments ... demonstrate the superiority' thus overstates what the main-text results show even under the preferred protocol.
  3. [Section III.C; Section III.B] The asymptotic reduction to WBCE is mathematically correct but is a weaker form of theoretical support than the paper claims. When r covers the whole image, the construction simply recovers the binary partition by setting Y_B = Y^- and Y_T = empty, and the reduction holds only after additionally choosing B_E=1 and B_B=lambda, with B_T unconstrained because the corresponding sum is empty. The limit therefore does not explain why the specific choices r=7, B_B=0.8, B_T=0.5 should improve edge precision; it establishes consistency with WBCE, not a predictive reason for the loss's effectiveness.
  4. [Section IV.A.1; Section II.D] The evaluation pipeline depends on the author's own unpublished preprints in ways that are difficult to verify independently. EES3 is the author's model [7], the dataset splits and preprocessing follow [7], and the noiseless-sample augmentation follows [19], both of which are unpublished preprints by the same author. The paper should either release the full training and evaluation code, exact split lists, and trained checkpoints, or anchor the evaluation to publicly available published protocols and splits. Without this, the claimed cross-model generality under a custom protocol cannot be assessed by a third party.
minor comments (4)
  1. [Section III.C, Eq. (7)] In the limit r going to infinity, Y_T is the empty set, so the texture term in Eq. (7) is formally a sum over no pixels; the text should state explicitly that empty-set terms contribute zero to avoid ambiguity about the role of B_T and W_T.
  2. [Table IV caption] The caption says 'The 1-pixel error tolerance is with regard to the test image with the lowest resolution,' which conflicts with the fixed 1-pixel radius stated in Section IV.A.1; please clarify how the tolerance is computed for UDED.
  3. [Section IV.A.1; Footnote 1] Footnote 1 contains a typo ('tansks' for 'tasks'), and the phrase 'in line with recent practices [19]' cites an unpublished preprint by the author; if a published source for the strict protocol exists, it should be cited instead.
  4. [Section IV.B.2; Figures 1-2] The 'perceptually superior' claim is supported only by two qualitative figure panels; no perceptual metric, user study, or quantitative sharpness measure is reported, so the perceptual claim is anecdotal.

Circularity Check

3 steps flagged · score 5.0 of 10

Trivialized WBCE-limit identity plus self-cited evaluation protocol make the central superiority claim only as strong as the author's own benchmark.

  1. self definitional [Section III-C (Asymptotic Behavior of EBT Loss), Eqs. (5)-(7)]
    "Therefore, by additionally setting BE = 1, BB = λ, the EBT loss reduces to: ... = − P i∈Y+ αlog( ˆyi) − P i∈Y− λ(1 − α)log(1 − ˆyi) / |Y| = LW BCE( ˆY , Y) (r → +∞)"

    The claimed theorem that EBT generalizes WBCE is achieved by exploiting free coefficients already present in the EBT definition. As r→∞ the masks satisfy YE=Y+, YB=Y−, YT=∅, and the adaptive weights collapse to WE=α, WB=1−α; choosing BE=1, BB=λ then forces the EBT formula to coincide with WBCE. The 'limit case' is therefore an identity by construction, not a derivation that constrains or predicts edge-detection behavior. It restates that WBCE lies inside EBT's parameter family and provides no independent theoretical evidence for the loss's superiority.

  2. self citation load bearing [Section IV-A-1 (Evaluation Scope and Model Setup); Section IV-A-2 (Dataset Preparation and Augmentation Strategy)]
    "Evaluation is based on the standard ODS, OIS, and AP metrics [22] under the strictest 1-pixel error tolerance without post-processing techniques such as NMS to precisely highlight the intrinsic capabilities of the models and loss functions, in line with recent practices [19]. ... Integration of noiseless samples into training, as suggested in [19]."

    The main-text claim that EBT is superior 'both quantitatively and perceptually' is established only under the 1-pixel, no-NMS protocol. Reference [19] is the same author's unpublished preprint that proposes this protocol, and EES3 is also the author's model (ref. [7]). When the standard NMS/relaxed-tolerance protocol is used, the paper's own appendix shows EBT frequently loses (e.g., NYUD2 AP: HED 0.648 vs 0.694, BDCN 0.683 vs 0.712, Dexi 0.687 vs 0.728, EES3 0.719 vs 0.730). Thus the central quantitative result is gated by a self-endorsed evaluation convention rather than an independent benchmark; removing the self-cited protocol removes the advertised superiority.

1 more flagged steps
  1. ansatz smuggled in via citation [Section II-D (Evaluation Protocols) and Section IV-A-1]
    "More recent studies advocate for standardized, stricter matching protocols [19], such as using a fixed 1-pixel radius, to ensure consistency and reduce error effects. Nonetheless, as a newly presented suggestion, this protocol has only been adopted in the most recent works."

    The paper refers to 'more recent studies' and 'recent practices' with citation [19], but [19] is the author's own prior preprint. The paper presents its own suggestion as if it were an external consensus, importing the 1-pixel/no-NMS evaluation ansatz via self-citation. This protocol is the exact gate that makes the EBT-vs-WBCE comparison favorable in the main text; the appendix under the traditional benchmark contradicts the abstract's unqualified superiority claim.

full rationale

The derivation chain contains no fabricated numbers and the EBT loss formula is an original contribution, so this is not a fully circular paper. However, two load-bearing steps weaken the claimed derivation. First, the paper's theoretical contribution—that EBT generalizes WBCE—is true but trivial: the limit-case reduction is obtained by setting the free balancing coefficients and weights in the EBT definition to reproduce WBCE exactly; it is an identity after construction rather than an independent theorem that supports the empirical claims. Second, the central experimental claim of superiority is evaluated only under a 1-pixel no-NMS protocol that the same author proposed in prior preprint [19], with the EES3 model, dataset splits, and noiseless-augmentation choices also coming from the author's own preprints [7, 19]. The paper's own appendix, using the traditional NMS/relaxed-tolerance benchmarks, shows EBT losing on NYUD2 for all four models and on several BIPED2 and UDED entries. This means the abstract-level assertion of general superiority is not independently established; it holds only when the self-cited protocol is accepted. The external baselines HED, BDCN, and Dexi provide some independence, which prevents a higher score, but the combination of a definitional identity and a self-cited evaluational gate justifies a moderate circularity score of 5.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method introduces no physical entities, but it does rely on three free hyperparameters (r=7, BB=0.8, BT=0.5), the boundary-band semantic assumption, and a nonstandard evaluation protocol. The protocol assumption is the most consequential: it is the main reason the paper's central claim appears to hold.

free parameters (3)
  • boundary radius r = 7
    Section III.B: 'r = 7 is set as the default, chosen by experiments.' It defines which non-edge pixels are treated as boundary pixels.
  • boundary weight BB = 0.8
    Section IV.A: balancing hyperparameters are fixed at BE=1.0, BB=0.8, BT=0.5. BB controls the weight of boundary pixels.
  • texture weight BT = 0.5
    Section IV.A: BT=0.5 is fixed. Only the ratios among BE, BB, BT matter for the loss.
assumptions (4)
  • ad hoc to paper Boundary pixels, defined as non-edge pixels within a (2r+1)-by-(2r+1) window around edge pixels, form a semantically useful supervision class.
    This is the core modeling assumption introduced in Section III.B. If the boundary band is not semantically meaningful, the method loses its rationale.
  • domain assumption The 1-pixel error tolerance without non-maximum suppression is the appropriate measure of edge detection quality.
    The main experiments use this protocol, which the author's prior work [19] advocates. The appendix shows results reverse under standard NMS protocols.
  • domain assumption The WBCE baselines and the pretrained EES3 extractors are faithfully reproduced from official implementations and from the author's prior work [7].
    The experiments compare against 'official implementations' and EES3 from [7]; if those are not faithfully reproduced, the comparisons are invalid.
  • domain assumption Ground-truth edge labels in all five datasets are reliable enough to define a stable tri-class decomposition.
    The paper itself discusses annotation noise in related work (Section II.A) but does not analyze how label noise affects the boundary/texture classes.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Edge-Boundary-Texture Loss: A Tri-Class Generalization of Weighted Binary Cross-Entropy for Enhanced Edge Detection." pith.science (2026). https://pith.science/paper/ECUFPTAI

@misc{pith2026250706569,
  author       = {Pith},
  title        = {Pith review of: Edge-Boundary-Texture Loss: A Tri-Class Generalization of Weighted Binary Cross-Entropy for Enhanced Edge Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ECUFPTAI}},
  note         = {Machine review of arXiv:2507.06569}
}
read the original abstract

Edge detection (ED) remains a fundamental task in computer vision, yet its performance is often hindered by the ambiguous nature of non-edge pixels near object boundaries. The widely adopted Weighted Binary Cross-Entropy (WBCE) loss treats all non-edge pixels uniformly, overlooking the structural nuances around edges and often resulting in blurred predictions. In this paper, we propose the Edge-Boundary-Texture (EBT) loss, a novel objective that explicitly divides pixels into three categories, edge, boundary, and texture, and assigns each a distinct supervisory weight. This tri-class formulation enables more structured learning by guiding the model to focus on both edge precision and contextual boundary localization. We theoretically show that the EBT loss generalizes the WBCE loss, with the latter becoming a limit case. Extensive experiments across multiple benchmarks demonstrate the superiority of the EBT loss both quantitatively and perceptually. Furthermore, the consistent use of unified hyperparameters across all models and datasets, along with robustness to their moderate variations, indicates that the EBT loss requires minimal fine-tuning and is easily deployable in practice.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 38 canonical work pages

  1. [19]

    Enhancing edge detection by texture handling architecture and noiseless training data

    Hao Shu. Enhancing edge detection by texture handling architecture and noiseless training data. 2025

    work page 2025

  2. [7]

    Boosting edge detection with pixel-wise feature selection: The extractor-selector paradigm, 2025

    Hao Shu. Boosting edge detection with pixel-wise feature selection: The extractor-selector paradigm, 2025

    work page 2025

  3. [1]

    Nazeri, E

    K. Nazeri, E. Ng, T. Joseph, F. Qureshi, and M. Ebrahimi. Edgeconnect: Structure guided image inpainting using edge prediction. In 2019 IEEE/CVF International Con- ference on Computer Vision Workshop (ICCVW) , pages 3265–3274, 2019

    work page 2019

  4. [2]

    C. H. Zhan, X. H. Duan, S. Y . Xu, Z. Song, and M. Luo. An improved moving object detection algorithm based on frame difference and edge detection. In Fourth international conference on image and graphics (ICIG 2007), pages 519–523. IEEE, 2007

    work page 2007

  5. [3]

    Muthukrishnan and M

    R. Muthukrishnan and M. Radha. Edge detection tech- niques for image segmentation. International Journal of Computer Science & Information Technology , 3(6):259, 2011

    work page 2011

  6. [4]

    S. N. Xie and Z. W. Tu. Holistically-nested edge detection. In Proceedings of the IEEE international conference on computer vision , pages 1395–1403, 2015

    work page 2015

  7. [5]

    J. Z. He, S. L. Zhang, M. Yang, Y . H. Shan, and T. J. Huang. Bdcn: Bi-directional cascade network for perceptual edge detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):100–113, 2022

    work page 2022

  8. [6]

    Soria, A

    X. Soria, A. Sappa, P. Humanante, and A. Akbarinia. Dense extreme inception network for edge detection. Pattern Recognition, 139:109461, 2023

    work page 2023

Show all 38 references
  1. [8]

    R. X. Deng, C. H. Shen, S. J. Liu, H. B. Wang, and X. R. Liu. Learning to predict crisp boundaries. In Com- puter Vision – ECCV 2018: 15th European Conference, Munich, Germany, September 8–14, 2018, Proceedings, Part VI, page 570–586. Springer-Verlag, 2018

    2018

  2. [9]

    Cetinkaya, S

    B. Cetinkaya, S. Kalkan, and E. Akbas. Ranked: Ad- dressing imbalance and uncertainty in edge detection using ranking-based losses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3239–3249, 2024

    2024

  3. [10]

    Martin, C

    D. Martin, C. Fowlkes, D. Tal, and J. Malik. A database of human segmented natural images and its application 9 to evaluating segmentation algorithms and measuring ecological statistics. In Proc. 8th Int’l Conf. Computer Vision, volume 2, pages 416–423, July 2001

    2001

  4. [11]

    D. A. M ´ely, J. k. Kim, M. McGill, Y . L. Guo, and T. Serre. A systematic comparison between visual cues for boundary detection. Vision Research, 120:93–107,

  5. [12]

    N. N. Silberman, D. Hoiem, P. P. Kohli, and R. Fergus. Indoor segmentation and support inference from rgbd images. In European Conference on Computer Vision , 2012

    2012

  6. [13]

    Everingham, L

    M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The pascal visual object classes (voc) challenge. International Journal of Computer Vision , 88(2):303–338, jun 2010

    2010

  7. [14]

    Soria, E

    X. Soria, E. Riba, and A. Sappa. Dense extreme inception network: Towards a robust cnn model for edge detec- tion. In 2020 IEEE Winter Conference on Applications of Computer Vision (WACV) , pages 1912–1921. IEEE Computer Society, mar 2020

    2020

  8. [15]

    M. Pu, Y . Huang, Q. Guan, and H. Ling. Rindnet: Edge detection for discontinuity in reflectance, illumination, normal and depth. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV) , pages 6859–

    2021

  9. [16]

    Soria, Y

    X. Soria, Y . Li, M. Rouhani, and A. D. Sappa. Tiny and efficient model for the edge detection generalization. In 2023 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW) , pages 1356–1365. IEEE Computer Society, oct 2023

    2023

  10. [17]

    Y . B. Fu and X. J. Guo. Practical edge detection via robust collaborative learning. In Proceedings of the 31st ACM International Conference on Multimedia , MM ’23, page 2526–2534. Association for Computing Machinery, 2023

    2023

  11. [18]

    C. Wang, D. Dai, S. Xia, Y . Liu, and G. Wang. One-stage deep edge detection based on dense-scale feature fusion and pixel-level imbalance learning. IEEE Transactions on Artificial Intelligence , 5(01):70–79, jan 2024

    2024

  12. [20]

    J. Kittler. On the accuracy of the sobel edge detector. Image and Vision Computing , 1(1):37–42, 1983

    1983

  13. [21]

    J. Canny. A computational approach to edge detection. IEEE Transactions on pattern analysis and machine intelligence, PAMI-8(6):679–698, 1986

    1986

  14. [22]

    Martin, C.C

    D.R. Martin, C.C. Fowlkes, and J. Malik. Learning to detect natural image boundaries using local brightness, color, and texture cues. IEEE Transactions on Pattern Analysis and Machine Intelligence, 26(5):530–549, 2004

    2004

  15. [23]

    J. J. Lim, C. L. Zitnick, and P. Doll ´ar. Sketch tokens: A learned mid-level representation for contour and object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 3158– 3165, 2013

    2013

  16. [24]

    Arbel ´aez, M

    P. Arbel ´aez, M. Maire, C. Fowlkes, and J. Malik. Contour detection and hierarchical image segmentation. IEEE Transactions on Pattern Analysis and Machine Intelli- gence, 33(5):898–916, 2011

    2011

  17. [25]

    Doll ´ar and C

    P. Doll ´ar and C. L. Zitnick. Fast edge detection using structured forests. IEEE Transactions on Pattern Analysis and Machine Intelligence , 37(8):1558–1570, 2015

    2015

  18. [26]

    X. F. Ren. Multi-scale improves boundary detection in natural images. In Computer Vision–ECCV 2008: 10th European Conference on Computer Vision, Marseille, France, October 12-18, 2008, Proceedings, Part III 10 , pages 533–545. Springer, 2008

    2008

  19. [27]

    Y . Liu, M. M. Cheng, X. W. Hu, K. Wang, and X. Bai. Richer convolutional features for edge detection. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 3000–3009, 2017

    2017

  20. [28]

    Z. Su, W. Z. Liu, Z. T. Yu, D. W. Hu, Q. Liao, Q. Tian, M. Pietik ¨ainen, and L. Liu. Pixel difference networks for efficient edge detection. In Proceedings of the IEEE/CVF international conference on computer vision , pages 5117–5127, 2021

    2021

  21. [29]

    M. Y . Pu, Y . P. Huang, Y . M. Liu, Q. J. Guan, and H. B. Ling. Edter: Edge detection with transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 1402–1412, 2022

    2022

  22. [30]

    J. H. Jie, Y . Guo, G. X. Wu, J. M. Wu, and B. J. Hua. Edgenat: Transformer for efficient edge detection. arXiv preprint arXiv:2408.10527, 2024

    2024 arXiv

  23. [31]

    C. X. Zhou, Y . P. Huang, M. C. Xiang, J. H. Ren, H. B. Ling, and J. Zhang. Generative edge detection with stable diffusion. arXiv preprint arXiv:2410.03080 , 2024

    2024 arXiv

  24. [32]

    Y . F. Ye, K. Xu, Y . H. Huang, R. J. Yi, and Z. P. Cai. Diffusionedge: Diffusion probabilistic model for crisp edge detection. In Proceedings of the AAAI conference on artificial intelligence , volume 38, pages 6675–6683, 2024

    2024

  25. [33]

    L. X. Huan, N. Xue, X. W. Zheng, W. He, J. Y . Gong, and G. S. Xia. Unmixing convolutional features for crisp edge detection. IEEE Trans. Pattern Anal. Mach. Intell. , 44(10 Part 1):6602–6609, oct 2022

    2022

  26. [34]

    K. Chen, J. G. Li, W. Y . Lin, J. See, J. Wang, L. Y . Duan, Z. B. Chen, C. W. He, and J. N. Zou. Towards accurate one-stage object detection with ap-loss. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5119–5127, 2019

    2019

  27. [35]

    Rethinking edge detection through perceptual asymmetry: The swbce loss, 2025

    Hao Shu. Rethinking edge detection through perceptual asymmetry: The swbce loss, 2025

    2025

  28. [36]

    Binarization-aware adjuster: Bridging continu- ous optimization and binary inference in edge detection, 2025

    Hao Shu. Binarization-aware adjuster: Bridging continu- ous optimization and binary inference in edge detection, 2025. 10 VII. A PPENDIX A. Evaluations Following Traditional Benchmarks Tables VI to X present the evaluation results under tradi- tional relaxed benchmarks, where ...

    2025

  29. [2016]

    Vision and the Statistics of the Natural Environ- ment

  30. [6868]

    IEEE Computer Society, oct 2021

    2021

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.