Pith. sign in

REVIEW 3 major objections 5 minor 115 references

PEdger++: Practical Edge Detection via Assembling Cross Information

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper proposes PEdger++, a collaborative learning framework that assembles cross-information from heterogeneous network architectures, different training moments, and multiple parameter samplings to make lightweight edge detection both

desk verdict Solid empirical engineering with a shaky theoretical wrapper: the accuracy/speed numbers are worth attention, but the claimed optimality of the ensemble weights does not hold up. read the letter →

arxiv 2508.11961 v1 pith:JAIBTKIP submitted 2025-08-16 cs.CV

Yuanbin Fu , Liang Li , Xiaojie Guo This is my paper
classification cs.CV
keywords edgedetectioncollaborativelearningBayesianneuralnetworksmodelensemblesofttargetsmomentumlightweightarchitectureepistemicuncertainty
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

PEdger++ claims that ensemble diversity can substitute for heavy architectures in edge detection. It trains a recurrent and a non-recurrent network together, fuses parameters across epochs with a momentum network, and approximates a Bayesian posterior over network parameters by sampling several parameter sets per epoch and weighting each by its validation-set generalization. The result, the paper reports, is a 716K-parameter detector trained from scratch that reaches ODS-F 0.835 at 92 FPS on BSDS with VOC data, and a 12.7M ResNet50-backed version that reaches ODS-F 0.857 at 37.6 FPS. If true, this would make high-accuracy edge detection practical on resource-constrained devices.

What carries the argument

The load-bearing object is the weighted parameter ensemble that stands in for a Bayesian posterior. At each epoch, S parameter samples are drawn from the momentum networks, each is assigned a pixel-wise weight by the closed-form formulas in Eqs. (8)–(9), and the fused prediction is a confidence-weighted combination of the recurrent and non-recurrent momentum-network outputs (Eq. 12). That fused map is blended with ground truth into soft targets that supervise training (Eq. 13), while the final inference parameters are a norm-weighted average of the samples (Eq. 10). The mechanism converts ensemble diversity into both a better training signal and a better test-time model without extra inferen

What would settle it

Compute the weights W_s from Eqs. (8)–(9) on a held-out validation split and check whether they sum to one and whether they yield a lower validation binary cross-entropy than equal weights or a directly optimized constrained solution; if equal weights or a numerical optimization beats the closed-form on ODS-F, the claimed optimality of the weight formulas is falsified.

Watch

Extended reading notes

Core claim

The central claim is that cross-information from three sources—heterogeneous architectures, training moments, and parameter samplings—can be assembled so that a compact, fast network matches or exceeds the accuracy of much larger detectors. The paper treats network parameters as random variables drawn from an approximate Bayesian posterior, samples S parameter sets per epoch via Monte Carlo dropout, and assigns each sample a pixel-wise ensemble weight derived from a claimed closed-form solution to a validation-loss minimization (Eqs. 8–9). The weighted predictions of the two momentum networks are fused by a confidence-aware rule (Eq. 12) and blended into soft targets (Eq. 13) that supervise

Load-bearing premise

The load-bearing premise is that the closed-form ensemble weights in Eqs. (8)–(9) actually solve the constrained validation-loss minimization in Eq. (6); the derivation treats each weight as if the fused prediction M did not depend on all the other weights and never enforces the stated sum-to-one constraint, so if those weights are not truly optimal, the Bayesian-posterior approximation and the soft-target refinement built on them lack their stated justification.

Editorial extensions

If this is right

  • A 716K-parameter edge detector trained from scratch can reach ODS-F 0.835 at 92 FPS, making state-of-the-art edge detection feasible on mobile and embedded hardware.
  • With a ResNet50 backbone the same framework reaches ODS-F 0.857 at 37.6 FPS with 12.7M parameters, beating recent Transformer- and diffusion-based detectors at a small fraction of their compute.
  • The validation-set-weighted parameter ensemble improves accuracy over other Bayesian-approximation strategies within the same framework, as shown in the paper's comparisons.
  • The efficient training variant (Algorithm 2) cuts training time by about 40% with only a 0.3–0.6 percentage-point drop in ODS-F, so the approach scales to larger data regimes.
  • The same architecture transfers to NYUD and Multicue with state-of-the-art-level results, suggesting the collaborative-learning principle is not dataset-specific.

Reading between the lines

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

  • The paper's gains may come more from ensemble diversity and soft-target smoothing than from the specific Bayesian-posterior interpretation; an ablation that replaces Eqs. (8)–(9) with equal weights would separate those effects.
  • The validation-weighted sampling scheme could transfer to other pixel-labeling tasks, such as semantic segmentation or depth estimation, where epistemic uncertainty is also a bottleneck, though the paper only tests edge detection.
  • Algorithm 2's use of stochastic weight pruning on a single network hints that the recurrent architecture's contribution may be mostly a source of diversity, not a necessary component; testing Algorithm 2 with the same compute budget could confirm this.
  • The reported sweet spot of S=3 parameter samples and Nv=30% validation split is empirical; the optimal values likely shift with dataset size and backbone capacity, so the method may need re-tuning for new domains.
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

3 major / 5 minor

Summary. This paper presents PEdger++, an edge detection framework that combines three forms of cross-information: a recurrent and a non-recurrent network trained collaboratively, momentum networks that fuse parameters across epochs, and multiple Monte-Carlo-style parameter samples whose predictions are merged with weights derived from a validation split. The final inference uses only the non-recurrent network, whose parameters are a weighted average of the sampled parameters. On BSDS500 with PASCAL VOC augmentation, the authors report ODS-F 0.835 (716K params, 92 FPS) without pre-training and 0.857 (12.7M params, 37.6 FPS) with a ResNet50 backbone, claiming state-of-the-art accuracy for efficient edge detection. The paper includes extensive comparisons on BSDS/NYUD/Multicue, ablations of each ensemble component, and an efficient training variant.

Significance. Assuming the empirical measurements are reproducible, the paper makes a practically important contribution: it demonstrates that a compact detector can reach accuracy comparable to much larger transformer/diffusion models while running one to two orders of magnitude faster. The ablation studies (Tabs. 9-11) support the claim that each of the three integration mechanisms contributes. The parameter study (Tab. 12) and scalability table (Tab. 7) are useful. The release of code is a strength. However, the theoretical apparatus in §3.3 — the claimed optimal weights and the mutual-information theorem — is not valid as written, and this is the main novelty over the authors' previous PEdger. The empirical results may stand, but the paper currently overstates its theoretical grounding.

major comments (3)
  1. [§3.3, Eqs. (6)-(9)] The closed-form weights in Eqs. (8)-(9) do not solve the stated constrained optimization. The objective in (6) depends on W_1,...,W_S through M = Σ_s W_s ∘ G_s. The Lagrange condition for an interior optimum of the BCE loss subject to Σ_s W_s = 1 is G_s(X;Θ_s) ∘ (M - Y)/(M(1-M)) = const for all s, which is not what Eq. (8) expresses. Eq. (8) is essentially a one-step negative-gradient update that treats M as fixed when differentiating; it omits the dependence of M on every W_s. It also does not enforce the constraints: at a non-edge pixel (Y=0) the bracketed term equals -G_s/(1-M) < 0, so every W_s is negative, and the normalization constant C cannot repair this. Because Algorithm 1 uses these weights to construct the soft targets (11)-(13) and the final parameter average (10), the "optimal" and "Bayesian posterior" claims are unsupported. The method should either be derived correctly (e
  2. [§3.3, Theorem 1] The proof of Theorem 1 is a definitional restatement rather than a substantive equivalence. Once p(Y|W,X) is defined by the product of Bernoulli probabilities with mean M=Σ_s W_s G_s, the conditional entropy is exactly the expected BCE, so minimizing the empirical loss is minimizing NLL by construction. The theorem does not show that the particular W_s from Eqs. (8)-(9) maximize mutual information, and the mutual-information objective I(Y;W_s|X) is not even well-defined without a distribution on W_s. The claim following Eqs. (8)-(9) that the calculated weights "can also maximize mutual information" is therefore unsupported. I recommend removing the theorem or replacing it with a precise statement about NLL minimization under the ensemble model.
  3. [§3.3 and Algorithm 1, sampling mechanism] Algorithm 1 (line 20) says "Sample parameters Θ_s" but the paper never specifies the distribution for the main (recurrent/non-recurrent) algorithm. The text invokes Monte Carlo Dropout, yet the architecture description in Sec. 3.1 does not mention dropout layers; the efficient variant (Algorithm 2) uses random weight pruning instead. This ambiguity makes the "multiple parameter samplings" component and the Bayesian interpretation non-reproducible. Please specify exactly how Θ_s^R and Θ_s^NR are generated, or revise the text to describe the actual stochastic weight perturbation.
minor comments (5)
  1. [Eq. (15)] The last norm has an unbalanced parenthesis: \|(1 - \tilde{Y}_{nt}) \circ \log(1 - G(X_{nt})\|_1 should be \|(1-\tilde{Y}_{nt}) \circ \log(1 - G(X_{nt}))\|_1.
  2. [References] References [3] and [4] are identical (Balabanov et al.), as are [51] and [52] (Lakshminarayanan et al.); please merge the duplicates.
  3. [Tabs. 1-2] The pre-trained configurations are described in the text as using two non-recurrent networks, but Algorithm 1 formulates recurrent/non-recurrent training. Please clarify which algorithm and network pair corresponds to each row.
  4. [Eq. (12)] The denominator can be zero when both predictions are exactly 0.5; add a small epsilon for numerical stability.
  5. [Table 3] The dataset column repeats 'BSDS' in a way that is confusing; a single multi-row header would be clearer.

Circularity Check

2 steps flagged · score 4.0 of 10

The optimal-weight 'derivation' in Eqs. (7)-(9) is self-referential, and Theorem 1 is a tautological restatement; the reported accuracy numbers, however, are external benchmark measurements, so the central empirical claim is not circular.

  1. self definitional [Section 3.3, Eqs. (7)-(9); Algorithm 1, lines 20-21]
    "The optimal weight of the s-th parameter sampling can be determined by the Lagrange multiplier method as: W_R_s := C ∘ Σ_{nv=1}^{N_v} G_R_m(X_nv; Θ_R_s) ∘ ( Y_nv / M_R_nv − 1 − Y_nv / 1 − M_R_nv ), (8) ... M_R_nv := Σ_{s:=1}^S W_R_s ∘ G_R_m(X_nv; Θ_R_s), (7)"

    Equation (8) is presented as the closed-form solution to the constrained minimization in Eq. (6), but its right-hand side contains M_R_nv, which by Eq. (7) is a weighted sum of the same W_R_s being solved for. Thus W_R_s is defined implicitly in terms of itself: the 'solution' is a fixed-point equation, not a Lagrange-multiplier solution. No λ enforcing Σ_s W_s = 1 appears, and the formula is not the stationary condition of Eq. (6) because M depends on all W_s. Algorithm 1 then treats these weights as known when building soft targets (Eqs. 11-13) and the final parameter average (Eq. 10), so the theoretical optimality claim reduces to a self-referential definition rather than a derivation.

  2. renaming known result [Section 3.3, Theorem 1 and its proof]
    "Theorem 1 (Mutual Information Maximization) Given validation dataset D_v ... the following optimization objectives are equivalent: 1. Mutual Information Maximization: max_{W_s} I(Y_nv; W_s | X_nv) ... 2. Validation Loss Minimization: min_{W_s} 1/N_v Σ_{nv=1}^{N_v} L(Σ_{s=1}^S W_s G(X_nv; Θ_s), Y_nv)."

    The proof sets p(Y | W, X) to the product of Bernoulli probabilities with M = Σ_s W_s G(X; Θ_s), so H(Y | W, X) equals the expected binary cross-entropy loss by construction. Therefore objective 1 and objective 2 are the same functional expressed in two notations; the theorem is a tautology. It is then invoked to assert that the weights from Eqs. (8)-(9) 'maximize mutual information,' but since Eq. (8) is self-referential and does not actually solve Eq. (6), the theorem supplies no independent support for the optimality claim.

full rationale

The paper's central empirical results—ODS-F 0.835/0.857, throughput, parameter counts—are obtained by training on fixed splits and evaluating on external test sets (BSDS test, NYUD, Multicue), so they are not constructed from the validation-fitted weights and are not circular. The circularity is confined to the theoretical framing of the ensemble weights. Eq. (8)/(9) is not a valid Lagrange solution because the right-hand side depends on M, which by Eq. (7) depends on the same weights; this is a self-referential definition rather than a derivation. Theorem 1 is a tautological equivalence between mutual-information maximization and cross-entropy minimization under the paper's own likelihood model, and it does not independently justify the fitted weights. No load-bearing self-citation was found: the authors' prior PEdger is used only as a baseline/comparison, not to justify the core derivation. The flawed theoretical derivation may overstate the Bayesian-posterior interpretation, but it does not invalidate the independently measured benchmark numbers. Hence the circularity is partial and theoretical, warranting a moderate score.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

The method rests on a standard set of ensemble and self-distillation assumptions. The free parameters are mostly tuned on BSDS and transferred to other datasets; the main theoretical load is carried by the validity of Eqs. (8)-(9).

free parameters (8)
  • λ (loss balance) = 1.1 (BSDS/Multicue), 1.3 (NYUD)
    Balances positive/negative terms in BCE via Eq. 16; tuned per dataset, no sensitivity analysis reported.
  • η_J (soft target weight) = 0.8
    Controls maximum influence of fused predictions on refined targets in Eq. 13; chosen via parameter study Tab. 12.
  • μ (momentum coefficient) = 0.5
    EMA coefficient for momentum networks in Eq. 5; stated as empirically set, no ablation.
  • S (number of parameter samplings) = 3
    Number of MC-dropout-style parameter samples; tuned in Tab. 12.
  • T (recurrent steps / scales) = 5 recurrent, 4 non-recurrent
    Number of recurrent steps and scale levels; tuned in Tab. 12.
  • Nv fraction = 30% of training set
    Validation split size used for weight computation; tuned in Tab. 12.
  • threshold for binary edge maps = 0.2
    Applied to BSDS and Multicue GT binarization; part of evaluation pipeline.
  • peak learning rate / warm-up = 0.001, 4 epochs
    Optimization schedule; not ablated.
assumptions (5)
  • domain assumption Ensemble generalization error decomposes as E = Ebar - Abar (Eq. 1).
    Borrowed from Krogh & Vedelsby and Ueda & Nakano; assumes diversity measures are well-defined and beneficial.
  • domain assumption MC Dropout with Bernoulli parameter sampling approximates the Bayesian posterior p(Θ|D).
    Invoked in Sec. 3.3 following Gal & Ghahramani; the approximation quality is not verified for this architecture.
  • domain assumption Momentum network parameter averaging preserves useful knowledge across training moments.
    Used in Eq. 5; standard in self-supervised learning; no formal guarantee.
  • domain assumption Soft targets / self-distillation improve generalization over hard labels.
    Used in Eq. 13; cited from label-smoothing and distillation literature.
  • standard math Held-out validation split Dv is a valid proxy for generalization to the test distribution.
    Assumed when computing weights on Dv; standard train/validation split assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PEdger++: Practical Edge Detection via Assembling Cross Information." pith.science (2026). https://pith.science/paper/JAIBTKIP

@misc{pith2026250811961,
  author       = {Pith},
  title        = {Pith review of: PEdger++: Practical Edge Detection via Assembling Cross Information},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JAIBTKIP}},
  note         = {Machine review of arXiv:2508.11961}
}
read the original abstract

Edge detection serves as a critical foundation for numerous computer vision applications, including object detection, semantic segmentation, and image editing, by extracting essential structural cues that define object boundaries and salient edges. To be viable for broad deployment across devices with varying computational capacities, edge detectors shall balance high accuracy with low computational complexity. While deep learning has evidently improved accuracy, they often suffer from high computational costs, limiting their applicability on resource-constrained devices. This paper addresses the challenge of achieving that balance: \textit{i.e.}, {how to efficiently capture discriminative features without relying on large-size and sophisticated models}. We propose PEdger++, a collaborative learning framework designed to reduce computational costs and model sizes while improving edge detection accuracy. The core principle of our PEdger++ is that cross-information derived from heterogeneous architectures, diverse training moments, and multiple parameter samplings, is beneficial to enhance learning from an ensemble perspective. Extensive experimental results on the BSDS500, NYUD and Multicue datasets demonstrate the effectiveness of our approach, both quantitatively and qualitatively, showing clear improvements over existing methods. We also provide multiple versions of the model with varying computational requirements, highlighting PEdger++'s adaptability with respect to different resource constraints. Codes are accessible at https://github.com/ForawardStar/EdgeDetectionviaPEdgerPlus/.

Figures

Figures reproduced from arXiv: 2508.11961 by the authors.

Figure 1
Figure 1. Predictions by two different parameter samplings. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Quantitative comparisons on the BSDS dataset. A larger area of the formed triangle indicates better performance, [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Timeline of key developments in edge detection. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Recurrent and non-recurrent architectures. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Precision-Recall curves on the BSDS dataset. Our [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Visual comparisons between our models with and without pre-training. [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: Visual comparisons with competitors pre-trained on ImageNet. Our results are obtained by PEdger++ w/ ResNet50. [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: Visual comparisons with methods in the absence of pre-training. [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: Visual comparison between different collaborative [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

115 extracted references · 56 canonical work pages

  1. [1]

    In: CVPR, pp

    Acuna, D., Kar, A., Fidler, S.: Devil is in the edges: learning semantic boundaries from noisy annotations. In: CVPR, pp. 11075–11083 (2019) 20 Yuanbin Fu et al

    2019

  2. [2]

    TPAMI 33(5), 898– 916 (2011)

    Arbelaez, P., Maire, M., Fowlkes, C., Malik, J.: Contour detec- tion and hierarchical image segmentation. TPAMI 33(5), 898– 916 (2011). DOI 10.1109/TPAMI.2010.161

    doi:10.1109/tpami.2010.161 2011

  3. [4]

    In: CVPR, pp

    Balabanov, O., Mehlig, B., Linander, H.: Bayesian posterior ap- proximation with stochastic ensembles. In: CVPR, pp. 13701– 13711 (2023)

    2023

  4. [5]

    In: ICPR, pp

    Barbano, R., Zhang, C., Arridge, S., Jin, B.: Quantifying model uncertainty in inverse problems via bayesian deep gradient descent. In: ICPR, pp. 1392–1399 (2021). DOI 10.1109/ ICPR48806.2021.9412521

    arXiv 2021

  5. [6]

    In: ICML (2018)

    Belghazi, M.I., et al.: Mine: mutual information neural estima- tion. In: ICML (2018)

    2018

  6. [7]

    In: CVPR, pp

    Bertasius, G., Shi, J., Torresani, L.: Deepedge: A multi-scale bifurcated deep network for top-down contour detection. In: CVPR, pp. 4380–4389 (2015). DOI 10.1109/CVPR.2015. 7299067

    doi:10.1109/cvpr.2015 2015

  7. [9]

    Breiman, L.: Random forests. Mach. Learn. 45(1), 5–32 (2001)

    2001

  8. [10]

    In: ICML, vol

    Brock, A., De, S., Smith, S.L., Simonyan, K.: High-performance large-scale image recognition without normalization. In: ICML, vol. 139, pp. 1059–1071 (2021)

    2021

Show all 115 references
  1. [11]

    TPAMI 8(6), 679–698 (1986)

    Canny, J.F.: A computational approach to edge detection. TPAMI 8(6), 679–698 (1986). DOI 10.1109/TPAMI.1986.4767851

    1986

  2. [12]

    In: ICCV , pp

    Cheng, S., Gokhale, T., Yang, Y .: Adversarial bayesian augmen- tation for single-source domain generalization. In: ICCV , pp. 11366–11376 (2023)

    2023

  3. [13]

    In: CVPR, pp

    Cheng, Z., Gadelha, M., Maji, S., Sheldon, D.: A bayesian per- spective on the deep image prior. In: CVPR, pp. 5443–5451 (2019)

    2019

  4. [14]

    In: CVPR (2019)

    Cheng, Z., Gadelha, M., Maji, S., Sheldon, D.: A bayesian per- spective on the deep image prior. In: CVPR (2019)

    2019

  5. [15]

    In: IJCAI, pp

    Cho, Y .J., Manoel, A., Joshi, G., Sim, R., Dimitriadis, D.: Hetero- geneous ensemble knowledge transfer for training large models in federated learning. In: IJCAI, pp. 2881–2887 (2022). DOI 10.24963/ijcai.2022/399

    2022 doi

  6. [16]

    In: ICML, vol

    Daxberger, E.A., Nalisnick, E.T., Allingham, J.U., Antor ´an, J., Hern´andez-Lobato, J.M.: Bayesian deep learning via subnetwork inference. In: ICML, vol. 139, pp. 2510–2521 (2021)

    2021

  7. [17]

    In: CVPR, pp

    Deng, J., Dong, W., Socher, R., Li, L., Li, K., Fei-Fei, L.: Ima- genet: a large-scale hierarchical image database. In: CVPR, pp. 248–255 (2009). DOI 10.1109/CVPR.2009.5206848

    2009

  8. [18]

    In: ACM MM, pp

    Deng, R., Liu, S.: Deep structural contour detection. In: ACM MM, pp. 304–312 (2020). DOI 10.1145/3394171.3413750

    2020

  9. [19]

    In: ACM MM, pp

    Deng, R., Liu, S., Wang, J., Wang, H., Zhao, H., Zhang, X.: Learning to decode contextual information for efficient con- tour detection. In: ACM MM, pp. 4435–4443 (2021). DOI 10.1145/3474085.3475593

    2021

  10. [20]

    In: ECCV , vol

    Deng, R., Shen, C., Liu, S., Wang, H., Liu, X.: Learning to pre- dict crisp boundaries. In: ECCV , vol. 11210, pp. 570–586 (2018). DOI 10.1007/978-3-030-01231-1 \ 35

    2018 doi

  11. [21]

    In: CVPR, pp

    Deng, Z., Yang, X., Xu, S., Su, H., Zhu, J.: Libre: a practical bayesian approach to adversarial detection. In: CVPR, pp. 972– 982 (2021)

    2021

  12. [22]

    In: NeurIPS, vol

    Deng, Z., Zhou, F., Zhu, J.: Accelerated linearized laplace ap- proximation for bayesian deep learning. In: NeurIPS, vol. 35, pp. 2695–2708 (2022)

    2022

  13. [23]

    In: CVPR, pp

    Diaz, R., Marathe, A.: Soft labels for ordinal regression. In: CVPR, pp. 4738–4747 (2019). DOI 10.1109/CVPR.2019.00487

    2019

  14. [24]

    In: ICCV , pp

    Doll ´ar, P., Zitnick, C.L.: Structured forests for fast edge detec- tion. In: ICCV , pp. 1841–1848 (2013). DOI 10.1109/ICCV .2013. 231

    2013 doi

  15. [25]

    TPAMI 23(3), 291–296 (2001)

    Elder, J.H., Goldberg, R.M.: Image editing in the contour do- main. TPAMI 23(3), 291–296 (2001). DOI 10.1109/34.910881

    2001 doi

  16. [26]

    TPAMI 46(4), 2027–2040 (2024)

    Franchi, G., Bursuc, A., Aldea, E., Dubuisson, S., Bloch, I.: En- coding the latent posterior of bayesian neural networks for un- certainty quantification. TPAMI 46(4), 2027–2040 (2024). DOI 10.1109/TPAMI.2023.3328829

    2027

  17. [27]

    In: ACM MM, pp

    Fu, Y ., Guo, X.: Practical edge detection via robust collaborative learning. In: ACM MM, pp. 2526–2534 (2023)

    2023

  18. [28]

    In: ICML, vol

    Gal, Y ., Ghahramani, Z.: Dropout as a bayesian approximation: representing model uncertainty in deep learning. In: ICML, vol. 48, pp. 1050–1059 (2016)

    2016

  19. [29]

    In: ACCV , vol

    Ganin, Y ., Lempitsky, V .S.: Nˆ4 -fields: neural network nearest neighbor fields for image transforms. In: ACCV , vol. 9004, pp. 536–551 (2014)

    2014

  20. [30]

    In: NIPS, vol

    Graves, A.: Practical variational inference for neural networks. In: NIPS, vol. 24 (2011)

    2011

  21. [31]

    In: CVPR, pp

    Grzech, D., Azampour, M.F., Glocker, B., Schnabel, J.A., Navab, N., Kainz, B., Folgoc, L.L.: A variational bayesian method for similarity learning in non-rigid image registration. In: CVPR, pp. 119–128 (2022)

    2022

  22. [32]

    In: ICML, vol

    Guo, C., Pleiss, G., Sun, Y ., Weinberger, K.Q.: On calibration of modern neural networks. In: ICML, vol. 70, pp. 1321–1330 (2017)

    2017

  23. [33]

    In: CVPR, pp

    Gupta, S., Arbelaez, P., Malik, J.: Perceptual organization and recognition of indoor scenes from rgb-d images. In: CVPR, pp. 564–571 (2013). DOI 10.1109/CVPR.2013.79

    2013 doi

  24. [34]

    In: CVPR, pp

    Hallman, S., Fowlkes, C.C.: Oriented edge forests for bound- ary detection. In: CVPR, pp. 1732–1740 (2015). DOI 10.1109/ CVPR.2015.7298782

    2015

  25. [35]

    IEEE Trans

    Hansen, L.K., Salamon, P.: Neural network ensembles. IEEE Trans. Pattern Anal. Mach. Intell. 12(10), 993–1001 (1990)

    1990

  26. [36]

    In: CVPR, pp

    He, J., Zhang, S., Yang, M., Shan, Y ., Huang, T.: Bi-directional cascade network for perceptual edge detection. In: CVPR, pp. 3828–3837 (2019). DOI 10.1109/CVPR.2019.00395

    2019

  27. [37]

    arXiv:1112.5745 (2011)

    Houlsby, N., et al.: Bayesian active learning for classification and preference learning. arXiv:1112.5745 (2011)

    2011 arXiv

  28. [38]

    Hu, Y ., Chen, Y ., Li, X., Feng, J.: Dynamic feature fusion for semantic edge detection. In: S. Kraus (ed.) IJCAI, pp. 782–788 (2019)

    2019

  29. [39]

    In: ECCV , pp

    Hu, Z., Zhen, M., Bai, X., Fu, H., Tai, C.: Jsenet: joint se- mantic segmentation and edge detection network for 3d point clouds. In: ECCV , pp. 222–239. Springer (2020). DOI 10.1007/ 978-3-030-58565-5 \ 14

    2020

  30. [40]

    In: NeurIPS (2023)

    Huang, Z., Lam, H., Zhang, H.: Efficient uncertainty quantifica- tion and reduction for over-parameterized neural networks. In: NeurIPS (2023)

    2023

  31. [41]

    Jordan, M.I., Ghahramani, Z., Jaakkola, T.S., Saul, L.K.: An in- troduction to variational methods for graphical models. Mach. Learn. 37(2), 183–233 (1999)

    1999

  32. [42]

    TPAMI45(6), 7220–7238 (2023)

    Kan, S., He, Z., Cen, Y ., Li, Y ., Mladenovic, V ., He, Z.: Con- trastive bayesian analysis for deep metric learning. TPAMI45(6), 7220–7238 (2023). DOI 10.1109/TPAMI.2022.3221486

    2023

  33. [43]

    In: CVPR, pp

    Kar, A., Biswas, P.K.: Fast bayesian uncertainty estimation and reduction of batch normalized single image super-resolution net- work. In: CVPR, pp. 4957–4966 (2021)

    2021

  34. [44]

    5574–5584 (2017)

    Kendall, A., Gal, Y .: What uncertainties do we need in bayesian deep learning for computer vision? In: NIPS, pp. 5574–5584 (2017)

    2017

  35. [46]

    In: ICLR (2023)

    Kim, S., Park, S., Kim, K., Yang, E.: Scale-invariant bayesian neural networks with connectivity tangent kernel. In: ICLR (2023)

    2023

  36. [47]

    In: ICCV , pp

    Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W., Doll´ar, P., Girshick, R.B.: Segment anything. In: ICCV , pp. 3992–4003 (2023) PEdger++: Practical Edge Detection via Assembling Cross Information 21

    2023

  37. [48]

    AAAI 34(04), 4477–4484 (2020)

    Krishnan, R., Subedar, M., Tickoo, O.: Specifying weight priors in bayesian deep neural networks with empirical bayes. AAAI 34(04), 4477–4484 (2020)

    2020

  38. [49]

    In: NIPS, pp

    Krogh, A., Vedelsby, J.: Neural network ensembles, cross valida- tion, and active learning. In: NIPS, pp. 231–238 (1994)

    1994

  39. [50]

    In: NeurIPS, pp

    Kull, M., Perell ´o-Nieto, M., K ¨angsepp, M., de Menezes e Silva Filho, T., Song, H., Flach, P.A.: Beyond temperature scaling: obtaining well-calibrated multi-class probabilities with dirichlet calibration. In: NeurIPS, pp. 12295–12305 (2019)

    2019

  40. [52]

    In: NIPS, pp

    Lakshminarayanan, B., Pritzel, A., Blundell, C.: Simple and scal- able predictive uncertainty estimation using deep ensembles. In: NIPS, pp. 6402–6413 (2017)

    2017

  41. [53]

    NeurIPS (2017)

    Lakshminarayanan, B., Pritzel, A., Blundell, C.: Simple and scalable predictive uncertainty estimation using deep ensembles. NeurIPS (2017)

    2017

  42. [54]

    PR 143, 109810 (2023)

    Liang, H., Liu, R., Wang, Z., Ma, J., Tian, X.: Variational bayesian deep network for blind poisson denoising. PR 143, 109810 (2023). DOI https://doi.org/10.1016/j.patcog.2023. 109810

    2023 doi

  43. [55]

    In: AAAI, pp

    Lienen, J., H ¨ullermeier, E.: From label smoothing to label relax- ation. In: AAAI, pp. 8583–8591 (2021). DOI https://doi.org/10. 1609/aaai.v35i10.17041

    2021

  44. [56]

    In: CVPR, pp

    Liu, T., Cheng, J., Tan, S.: spectral bayesian uncertainty for im- age super-resolution. In: CVPR, pp. 18166–18175 (2023)

    2023

  45. [57]

    TPAMI 41(8), 1939–1946 (2019)

    Liu, Y ., Cheng, M.M., Hu, X., Bian, J.W., Zhang, L., Bai, X., Tang, J.: Richer convolutional features for edge detection. TPAMI 41(8), 1939–1946 (2019). DOI 10.1109/TPAMI.2018. 2878849

    1939 doi

  46. [58]

    In: CVPR, pp

    Liu, Y ., Lew, M.S.: Learning relaxed deep supervision for better edge detection. In: CVPR, pp. 231–240 (2016). DOI 10.1109/ CVPR.2016.32

    2016

  47. [59]

    In: AAAI, pp

    Liufu, X., Tan, C., Lin, X., Qi, Y ., Li, J., Hu, J.: Sauge: taming sam for uncertainty-aligned multi-granularity edge detection. In: AAAI, pp. 5766–5774 (2025)

    2025

  48. [60]

    In: CVPR, pp

    Lu, J., Cui, C., Ma, Y ., Bera, A., Wang, Z.: Quantifying uncer- tainty in motion prediction with variational bayesian mixture. In: CVPR, pp. 15428–15437 (2024)

    2024

  49. [61]

    In: IJCAI, pp

    Lu, Y ., He, W.: Selc: self-ensemble label correction improves learning with noisy labels. In: IJCAI, pp. 3278–3284 (2022). DOI 10.24963/ijcai.2022/455

    2022 doi

  50. [62]

    In: ECCV , pp

    Lukov, T., Zhao, N., Lee, G.H., Lim, S.: Teaching with soft label smoothing for mitigating noisy labels in facial expressions. In: ECCV , pp. 648–665 (2022). DOI 10.1007/978-3-031-19775-8\ 38

    2022 doi

  51. [63]

    Neural Comput

    MacKay, D.J.C.: A practical bayesian framework for backpropa- gation networks. Neural Comput. 4(3), 448–472 (1992)

    1992

  52. [64]

    In: NeurIPS, vol

    Maddox, W.J., Izmailov, P., Garipov, T., Vetrov, D.P., Wilson, A.G.: A simple baseline for bayesian uncertainty in deep learn- ing. In: NeurIPS, vol. 32 (2019)

    2019

  53. [65]

    In: NeurIPS, pp

    Malinin, A., Gales, M.J.F.: Predictive uncertainty estimation via prior networks. In: NeurIPS, pp. 7047–7058 (2018)

    2018

  54. [66]

    In: NeurIPS, pp

    Malinin, A., Gales, M.J.F.: Reverse kl-divergence training of prior networks: improved uncertainty and adversarial robustness. In: NeurIPS, pp. 14520–14531 (2019)

    2019

  55. [67]

    In: ECCV (2016)

    Maninis, K.K., Pont-Tuset, J., Arbel ´aez, P., Gool, L.V .: Convolu- tional oriented boundaries. In: ECCV (2016)

    2016

  56. [68]

    TPAMI 40(4), 819–833 (2017)

    Maninis, K.K., Pont-Tuset, J., Arbel ´aez, P., Van Gool, L.: Convo- lutional oriented boundaries: From image segmentation to high- level tasks. TPAMI 40(4), 819–833 (2017). DOI 10.1109/ TPAMI.2017.2700300

    2017

  57. [69]

    TPAMI 26(5), 530–549 (2004)

    Martin, D.R., Fowlkes, C.C., Malik, J.: Learning to detect natu- ral image boundaries using local brightness, color, and texture cues. TPAMI 26(5), 530–549 (2004). DOI 10.1109/TPAMI. 2004.1273918

    2004 arXiv

  58. [70]

    Scientific Reports 11(1) (2021)

    Mobiny, A., Yuan, P., Moulik, S.K., Garg, N., Nguyen, H.V .: Dropconnect is effective in modeling uncertainty of bayesian deep networks. Scientific Reports 11(1) (2021)

    2021

  59. [71]

    In: CVPR, pp

    Mottaghi, R., Chen, X., Liu, X., Cho, N.G., Lee, S.W., Fidler, S., Urtasun, R., Yuille, A.: The role of context for object detection and semantic segmentation in the wild. In: CVPR, pp. 891–898 (2014). DOI 10.1109/CVPR.2014.119

    2014 doi

  60. [72]

    Neal, R.M.: Bayesian learning for neural networks. Ph.D. thesis, University of Toronto, Canada (1995)

    1995

  61. [73]

    In: ICML, vol

    Ober, S.W., Aitchison, L.: Global inducing point variational pos- teriors for bayesian neural networks and deep gaussian processes. In: ICML, vol. 139, pp. 8248–8259 (2021)

    2021

  62. [74]

    In: NeurIPS, vol

    Osawa, K., Swaroop, S., Khan, M.E.E., Jain, A., Eschen- hagen, R., Turner, R.E., Yokota, R.: Practical deep learning with bayesian principles. In: NeurIPS, vol. 32 (2019)

    2019

  63. [75]

    PR 156, 110820 (2024)

    Pan, Z., Jiang, P., Zeng, Q., Li, G., Tu, C.: Category-agnostic semantic edge detection by measuring neural representation ran- domness. PR 156, 110820 (2024)

    2024

  64. [76]

    In: CVPR, pp

    Pu, M., Huang, Y ., Liu, Y ., Guan, Q., Ling, H.: EDTER:edge detection with transformer. In: CVPR, pp. 1392–1402 (2022). DOI 10.1109/CVPR52688.2022.00146

    2022

  65. [77]

    In: ICCV , pp

    Qu, C., Liu, W., Taylor, C.J.: Bayesian deep basis fitting for depth completion with uncertainty. In: ICCV , pp. 16147–16157 (2021)

    2021

  66. [78]

    TPAMI 39(6), 1137–1149 (2017)

    Ren, S., He, K., Girshick, R.B., Sun, J.: Faster r-cnn: to- wards real-time object detection with region proposal networks. TPAMI 39(6), 1137–1149 (2017). DOI 10.1109/TPAMI.2016. 2577031

    2017 doi

  67. [79]

    In: CVPR, pp

    Ren, Z., Shakhnarovich, G.: Image segmentation by cascaded re- gion agglomeration. In: CVPR, pp. 2011–2018 (2013). DOI 10.1109/CVPR.2013.262

    2011 doi

  68. [80]

    Picture Process- ing and

    S., P.J.M.: Object enhancement and extraction. Picture Process- ing and. Psychopictorics (1970)

    1970

  69. [81]

    In: NeurIPS (2023)

    Schweighofer, K., Aichberger, L., Ielanskyi, M., Klambauer, G., Hochreiter, S.: Quantification of uncertainty with adversarial models. In: NeurIPS (2023)

    2023

  70. [82]

    In: NeurIPS, vol

    Seligmann, F., Becker, P., V olpp, M., Neumann, G.: Beyond deep ensembles: a large-scale evaluation of bayesian deep learning un- der distribution shift. In: NeurIPS, vol. 36, pp. 29372–29405 (2023)

    2023

  71. [83]

    In: NeurIPS, pp

    Sensoy, M., Kaplan, L.M., Kandemir, M.: Evidential deep learn- ing to quantify classification uncertainty. In: NeurIPS, pp. 3183– 3193 (2018)

    2018

  72. [84]

    The Bell system technical journal 27(3), 379–423 (1948)

    Shannon, C.E.: A mathematical theory of communication. The Bell system technical journal 27(3), 379–423 (1948)

    1948

  73. [85]

    In: CVPR, pp

    Shen, W., Wang, X., Wang, Y ., Bai, X., Zhang, Z.: Deepcontour: A deep convolutional feature learned by positive-sharing loss for contour detection. In: CVPR, pp. 3982–3991 (2015). DOI 10. 1109/CVPR.2015.7299024

    2015

  74. [86]

    In: ICLR (2015)

    Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. In: ICLR (2015). DOI https: //doi.org/10.48550/arXiv.1409.1556

    2015 doi arXiv

  75. [87]

    Journal of neuroscience methods 26(1), 75–81 (1988)

    Smith Jr, T., Marks, W., Lange, G., Sheriff Jr, W., Neale, E.: Edge detection in images using marr-hildreth filtering tech- niques. Journal of neuroscience methods 26(1), 75–81 (1988)

    1988

  76. [88]

    Sobel, I.: Camera models and machine perception. Tech. rep. (1972)

    1972

  77. [89]

    Srivastava, N., Hinton, G.E., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: a simple way to prevent neural net- works from overfitting. J. Mach. Learn. Res. 15(1), 1929–1958 (2014)

    1929

  78. [91]

    TPAMI 45(12), 14956–14974 (2023) 22 Yuanbin Fu et al

    Su, Z., Zhang, J., Wang, L., Zhang, H., Liu, Z., Pietik ¨ainen, M., Liu, L.: Lightweight pixel difference networks for efficient visual representation learning. TPAMI 45(12), 14956–14974 (2023) 22 Yuanbin Fu et al

    2023

  79. [92]

    In: ICCV , pp

    Subedar, M., Krishnan, R., Lopez-Meyer, P., Tickoo, O., Huang, J.: Uncertainty-aware audiovisual activity recognition using deep bayesian variational inference. In: ICCV , pp. 6300–6309 (2019)

    2019

  80. [93]

    Teye, M., Azizpour, H., Smith, K.: Bayesian uncertainty es- timation for batch normalized deep networks. In: J.G. Dy, A. Krause (eds.) ICML, Proceedings of Machine Learning Re- search, vol. 80, pp. 4914–4923 (2018)

    2018

  81. [94]

    In: ICNN, pp

    Ueda, N., Nakano, R.: Generalization error of ensemble estima- tors. In: ICNN, pp. 90–95 (1996)

    1996

  82. [95]

    TIP 27(3), 1542–1553 (2018)

    Wang, D., Tan, X.: Robust distance metric learning via bayesian inference. TIP 27(3), 1542–1553 (2018)

    2018

  83. [96]

    In: CVPR, pp

    Wang, H., Joshi, D., Wang, S., Ji, Q.: Gradient-based uncertainty attribution for explainable bayesian deep learning. In: CVPR, pp. 12044–12053 (2023)

    2023

  84. [97]

    In: CVPR, pp

    Wang, J., Lukasiewicz, T.: Rethinking bayesian deep learning methods for semi-supervised volumetric medical image segmen- tation. In: CVPR, pp. 182–190 (2022)

    2022

  85. [98]

    In: IJCAI, pp

    Wang, K., Nie, Y ., Fang, C., Han, C., Wu, X., Wang, X., Lin, L., Zhou, F., Li, G.: Double-check soft teacher for semi-supervised object detection. In: IJCAI, pp. 1430–1436 (2022). DOI 10. 24963/ijcai.2022/199

    2022

  86. [99]

    In: CVPR, pp

    Wang, Y ., Zhao, X., Huang, K.: Deep crisp boundaries. In: CVPR, pp. 1724–1732 (2017). DOI 10.1109/CVPR.2017.187

    2017 doi

  87. [100]

    In: NeurIPS (2023)

    Warburg, F., Miani, M., Brack, S., Hauberg, S.: Bayesian met- ric learning for uncertainty quantification in image retrieval. In: NeurIPS (2023)

    2023

  88. [101]

    In: NeurIPS, vol

    Warburg, F., Miani, M., Brack, S., Hauberg, S.r.: Bayesian met- ric learning for uncertainty quantification in image retrieval. In: NeurIPS, vol. 36, pp. 69178–69190 (2023)

    2023

  89. [102]

    In: NeurIPS, vol

    Wild, V .D., Ghalebikesabi, S., Sejdinovic, D., Knoblauch, J.: A rigorous link between deep ensembles and (variational) bayesian methods. In: NeurIPS, vol. 36, pp. 39782–39811 (2023)

    2023

  90. [103]

    In: NeurIPS, vol

    Wild, V .D., Hu, R., Sejdinovic, D.: Generalized variational infer- ence in function spaces: gaussian measures meet bayesian deep learning. In: NeurIPS, vol. 35, pp. 3716–3730 (2022)

    2022

  91. [104]

    In: NeurIPS, vol

    Wilson, A.G., Izmailov, P.: Bayesian deep learning and a prob- abilistic perspective of generalization. In: NeurIPS, vol. 33, pp. 4697–4708 (2020)

    2020

  92. [105]

    In: ICML, vol

    Xiao, Z., Shen, J., Zhen, X., Shao, L., Snoek, C.: A bit more bayesian: domain-invariant learning with uncertainty. In: ICML, vol. 139, pp. 11351–11361 (2021)

    2021

  93. [106]

    IJCV 125(1- 3), 3–18 (2017)

    Xie, S., Tu, Z.: Holistically-nested edge detection. IJCV 125(1- 3), 3–18 (2017). DOI 10.1007/s11263-017-1004-z

    2017 doi

  94. [107]

    In: NIPS, pp

    Xu, D., Ouyang, W., Alameda-Pineda, X., Ricci, E., Wang, X., Sebe, N.: Learning deep structured multi-scale features using attention-gated crfs for contour prediction. In: NIPS, pp. 3961– 3970 (2017). DOI https://doi.org/10.48550/arXiv.1801.00524

    2017 doi arXiv

  95. [108]

    Neural Net- works 145, 248–259 (2022)

    Xuan, W., Huang, S., Liu, J., Du, B.: Fcl-net: towards accurate edge detection via fine-scale corrective learning. Neural Net- works 145, 248–259 (2022). DOI 10.1016/j.neunet.2021.10.022

    2022 doi

  96. [109]

    In: ACM MM, pp

    Xuan, W., Zhao, S., Yao, Y ., Liu, J., Liu, T., Chen, Y ., Du, B., Tao, D.: Pnt-edge: towards robust edge detection with noisy labels by learning pixel-level noise transitions. In: ACM MM, pp. 1924– 1932 (2023)

    1924

  97. [110]

    In: CVPR, pp

    Yang, J., Price, B.L., Cohen, S., Lee, H., Yang, M.: Object con- tour detection with a fully convolutional encoder-decoder net- work. In: CVPR, pp. 193–202 (2016). DOI 10.1109/CVPR. 2016.28

    2016 doi

  98. [111]

    Ye, Y ., Xu, K., Huang, Y ., Yi, R., Cai, Z.: Diffusionedge: dif- fusion probabilistic model for crisp edge detection. In: M.J. Wooldridge, J.G. Dy, S. Natarajan (eds.) AAAI, pp. 6675–6683 (2024)

    2024

  99. [112]

    In: CVPR, pp

    Yu, Z., Feng, C., Liu, M., Ramalingam, S.: Casenet: deep category-aware semantic edge detection. In: CVPR, pp. 1761– 1770 (2017)

    2017

  100. [113]

    In: ECCV , vol

    Yu, Z., Liu, W., Zou, Y ., Feng, C., Ramalingam, S., Kumar, B.V .K.V ., Kautz, J.: Simultaneous edge alignment and learning. In: ECCV , vol. 11207, pp. 400–417 (2018)

    2018

  101. [114]

    TPAMI 43(12), 4306–4322 (2021)

    Zhang, H., Chen, B., Cong, Y ., Guo, D., Liu, H., Zhou, M.: Deep autoencoding topic model with scalable hybrid bayesian infer- ence. TPAMI 43(12), 4306–4322 (2021). DOI 10.1109/TPAMI. 2020.3003660

    2021

  102. [115]

    In: CVPR, pp

    Zhao, R., Xu, W., Su, H., Ji, Q.: Bayesian hierarchical dynamic model for human action recognition. In: CVPR, pp. 7733–7742 (2019)

    2019

  103. [116]

    TPAMI 45(5), 5889–5903 (2023)

    Zheng, D., Zhang, X., Ma, K., Bao, C.: Learn from unpaired data for image restoration: a variational bayes approach. TPAMI 45(5), 5889–5903 (2023). DOI 10.1109/TPAMI.2022.3215571

    2023

  104. [117]

    In: CVPR, pp

    Zhou, C., Huang, Y ., Pu, M., Guan, Q., Deng, R., Ling, H.: Muge: multiple granularity edge detection. In: CVPR, pp. 25952–25962 (2024)

    2024

  105. [118]

    In: CVPR, pp

    Zhou, C., Huang, Y ., Pu, M., Guan, Q., Huang, L., Ling, H.: The treasure beneath multiple annotations: an uncertainty-aware edge detector. In: CVPR, pp. 15507–15517 (2023)

    2023

  106. [119]

    In: ICLR (2021)

    Zhou, H., Song, L., Chen, J., Zhou, Y ., Wang, G., Yuan, J., Zhang, Q.: Rethinking soft labels for knowledge distillation: a bias-variance tradeoff perspective. In: ICLR (2021). DOI https://doi.org/10.48550/arXiv.2102.00650

    2021 doi arXiv

  107. [120]

    In: ICCV , pp

    Zi, B., Zhao, S., Ma, X., Jiang, Y .: Revisiting adversarial robust- ness distillation: robust soft labels make student better. In: ICCV , pp. 16423–16432 (2021). DOI 10.1109/ICCV48922.2021.01613

    2021

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.