Co-PLNet: A Collaborative Point-Line Network for Prompt-Guided Wireframe Parsing

Chao Wang; Cheng Dai; Hao Qin; Jinglei Feng; Xuanying Li; Yuqi Ouyang; Yuxiang Luo

REVIEW 4 major objections 4 minor 18 references

Co-PLNet shows wireframe parsing accuracy and consistency improve when junction and line predictions prompt each other, while running at 76.8 FPS.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 08:01 UTC pith:QEABNW3R

load-bearing objection Co-PLNet is a credible architectural extension with ablations that look internally consistent, but the sAP metric in Eq. (11) is not the standard one, so the headline gains over HAWP/PLNet are not currently comparable. the 4 major comments →

arxiv 2601.18252 v2 pith:QEABNW3R submitted 2026-01-26 cs.CV cs.AIcs.LGstat.ML

Co-PLNet: A Collaborative Point-Line Network for Prompt-Guided Wireframe Parsing

Chao Wang , Xuanying Li , Cheng Dai , Jinglei Feng , Yuxiang Luo , Hao Qin , Yuqi Ouyang This is my paper

classification cs.CV cs.AIcs.LGstat.ML

keywords wireframe parsingline segment detectionjunction detectionpoint-line collaborationspatial promptssparse cross-attentionstructural average precisionreal-time efficiency

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Wireframe parsing recovers the line segments and junctions that make up the visible structure of a scene, and it matters for tasks like visual SLAM that need consistent geometry. The paper tries to establish that doing this jointly is better than the standard approach of detecting lines and junctions separately and reconciling them afterward. Its method, Co-PLNet, converts early junction and line detections into spatial prompts and exchanges those prompts between the two prediction paths, then refines line features with sparse gated cross-attention. Reported results on the Wireframe and YorkUrban datasets show accuracy gains at every structural-average-precision threshold (sAP, a matching-based accuracy measure for line segments) on both color and greyscale inputs, with endpoint mismatch roughly halved, at 76.8 frames per second. If the claims hold, the practical payoff is a real-time wireframe parser whose lines and junctions agree better, which is exactly what downstream geometric systems need.

Core claim

The central claim is that bidirectional conditioning between junction and line prediction produces more accurate and more consistent wireframes than independent prediction with post-hoc alignment. Co-PLNet first generates coarse junction coordinates and dense line-proposal maps, encodes them into low-channel spatial prompts, and then conditions each prediction task on the other task's prompt inside a cross-guidance decoder that uses windowed sparse multi-head attention and learned gating. On the Wireframe and YorkUrban datasets the paper reports consistently higher structural average precision at all thresholds for both color and greyscale input, and an endpoint-mismatch reduction from 12.4%

What carries the argument

The load-bearing mechanism is the pair of prompt modules: the Point-Line Prompt Encoder (PLP-Encoder) and the Cross-Guidance Line Decoder (CGL-Decoder). The encoder converts early junction heatmaps and per-pixel line-parameter fields into sparse junction coordinates and dense endpoint proposals, then maps them to compact spatially aligned prompt maps y_J^(0), y_L^(0). The decoder fuses those prompts with a shared feature map, applies sparse multi-head cross-attention within spatial windows so that line features attend to junction context and vice versa, and injects the attended features through learned gating masks. The formal idea is the factorization p(y|I)=sum_{(y_J,y_L)} p_J(y_J|I,y_L^(0

Load-bearing premise

The comparison assumes Eq. (11) is the same structural-average-precision metric prior work reports; as written it uses squared endpoint distances, and the paper also omits the detailed loss definitions (deferring them to a prior paper), so the headline numbers and the training objective are not fully self-contained.

What would settle it

Run the released code on the Wireframe test set and re-score all methods under the literal rule of Eq. (11) with l=5,10,15; if Co-PLNet's reported 68.4/72.3/73.8 does not reproduce, or if re-scoring the baselines under the same rule changes the ordering, the central accuracy claim fails. Independent checks: recompute the endpoint mismatch rate with the same junction-association threshold, and re-run the training with the loss definitions from the referenced paper to confirm the reported ablations.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

If the reported accuracy holds, wireframe parsers can be improved without a speed penalty, since Co-PLNet runs at 76.8 FPS while outperforming the compared baselines on both datasets.
Mutual prompting cuts endpoint mismatch from 12.4% to 9.6% with both prompts and to 7.8% with the full decoder on Wireframe, a direct measure of point-line consistency that downstream geometry systems would benefit from.
Sparse attention is what preserves the real-time claim: dense attention improves accuracy only marginally while dropping inference speed to 42.1 FPS.
The gains are reported on both the in-domain Wireframe test set and the cross-domain YorkUrban set, and for both color and grayscale inputs, suggesting the collaboration, not input color, drives the improvement.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same prompt-exchange design could be applied to other paired structured-output tasks, such as edge detection with junction/endpoint agreement or vanishing-point and line-group estimation, where two heads must be mutually consistent; the paper does not test these settings.
Because the matching rule in Eq. (11) uses squared endpoint distance, the labels sAP5/sAP10/sAP15 may not mean what they appear to mean; if the formula is literal, 'sAP5' is a tolerance of roughly 2.24 pixels, and all comparisons against published numbers would need re-scoring before the ranking can be trusted.
A natural next evaluation is to measure the impact on actual SLAM or 3D reconstruction rather than detection metrics, since the motivation is downstream geometry but the experiments stop at wireframe accuracy.
The training-loss definitions are omitted and deferred to a prior paper, so the exact objective being optimized is not fully specified; reproduction depends on retrieving those definitions and confirming the prompt modules do not require additional loss terms.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

Co-PLNet is a credible architectural extension with ablations that look internally consistent, but the sAP metric in Eq. (11) is not the standard one, so the headline gains over HAWP/PLNet are not currently comparable.

read the letter

Two things to know before reading: the architecture is a reasonable, if modest, extension of HAWP/PLNet that lets junction and line predictions prompt each other through a sparse cross-attention decoder; and the main accuracy numbers are probably not on the same scale as the published baselines because Eq. (11) defines sAP in an unusual way.

The actual contribution is the mutual-prompting scheme: early junction and line detections are encoded into small spatial prompt maps, and the cross-guidance line decoder uses windowed sparse attention to refine line features with those prompts. That is a genuine architectural move, even though most building blocks (SuperPoint, the U-Net, HAFM, LOI verification) come from prior work. The ablations in Table II are internally consistent and show each component adds something: point-to-line and line-to-point prompts both reduce the endpoint mismatch rate, and sparse attention maintains ~77 FPS versus 42 for dense attention. If I ignore the evaluation issues, the method looks plausible.

The soft spots are concentrated in the evaluation and are real, not manufactured. First, Eq. (11) defines a true positive as ||ĉ1−c1||² + ||ĉ2−c2||² ≤ l, whereas the wireframe-parsing literature (including the cited L-CNN and HAWPv2) uses max(||ĉ1−c1||, ||ĉ2−c2||) ≤ l. For l=5, Eq. (11) implies a ~2.24 px per-endpoint tolerance, so Table I's sAP5 is not measuring the same thing as the baseline numbers. If that is a typo, the evaluation code needs to be checked; if it is literal, the claim of consistent improvement across all thresholds is unsupported. This is load-bearing because every comparison in Tables I–III rests on it. Second, the window size for sparse attention is selected on the test sets (Table III), which inflates accuracy estimates. Third, no error bars or variance are reported. Fourth, the loss definitions are omitted and the paper refers to reference [2], which is the AIRSLAM paper rather than the PLNet source containing those losses. None of these are fatal on their own, but together they make the numerical results provisional.

Who should read this: anyone working on wireframe parsing or joint point-line perception will find the mutual-prompt idea interesting and the ablations instructive, but should not quote the sAP numbers until the metric and code are clarified. My recommendation: send it to peer review, but only if the authors fix Eq. (11), rerun with validation-based hyperparameter selection, report variance, and provide the actual loss definitions. If they can do that and the code matches, this is a solid incremental paper. If the metric issue turns out to be a literal definition rather than a typo, the headline gains likely collapse.

Referee Report

4 major / 4 minor

Summary. The paper proposes Co-PLNet, a wireframe-parsing network that performs early junction and line detection and then exchanges spatial prompts between the two tasks through a Point-Line Prompt Encoder and a Cross-Guidance Line Decoder with sparse multi-head cross-attention. The authors claim consistent improvements over prior methods such as HAWPv2 and PLNet on the Wireframe and YorkUrban datasets under structural average precision (sAP), while operating at 76.8 FPS. The paper includes ablations of the prompt, local fusion, sparse attention, and attention window size, and reports an endpoint-mismatch metric as evidence of improved point-line consistency.

Significance. If the central empirical claim is correct, Co-PLNet is a useful step toward tighter integration of junction and line predictions in wireframe parsing, and the proposed prompt-based cross-task conditioning is a plausible and interesting design. The explicit code link and the inclusion of both color and greyscale evaluations are strengths. However, the claim is currently not established because the evaluation metric defined in Eq. (11) may not be the same sAP used by the published baselines, and the training losses are not specified in the manuscript. The architectural contribution is tangible, but the quantitative comparison needs correction and verification before the claimed superiority can be accepted.

major comments (4)

[§III-A, Eq. (11)] The true-positive criterion is stated as ||ĉ1−c1||² + ||ĉ2−c2||² ≤ l with l∈{5,10,15}. In the wireframe literature (L-CNN [9], HAWPv2 [11]), sAP uses a per-endpoint Euclidean threshold, i.e., max(||ĉ1−c1||, ||ĉ2−c2||) ≤ l. Under Eq. (11), sAP5 corresponds to a 4D ball of radius √5 ≈ 2.24 px, which is considerably stricter than a 5-px per-endpoint tolerance. If Eq. (11) is literal, every sAP number in Tables I–III is on a different scale and cannot be compared with the published baselines. If it is a typo, the evaluation code must be checked to determine which criterion was actually used. This issue is load-bearing because the headline claim of consistent outperformance rests entirely on this metric.
[§III-E, Table III] The attention window size is selected by comparing test-set sAP values (w=4,8,16) and then choosing w=8 because it "peaks" on Wireframe and YorkUrban. This is test-set tuning, which risks overfitting to the evaluation set and invalidates the reported test numbers as unbiased estimates. The authors should either use a held-out validation split for parameter selection or clearly state that w=8 is chosen on the test set and treat the comparison as exploratory. Additionally, no error bars or multiple-run statistics are provided, so differences such as 67.9 vs. 68.1 in Table II may be within run-to-run variation. These issues affect the robustness of the ablation conclusions.
[§II-C, Eq. (10)] The total loss decomposes into L_line, L_junc, L_aux, and L_LOI, but the definitions of L_line, L_junc, and L_aux are omitted, with the text saying they "can be found in [2]." Reference [2] in the bibliography is AirSLAM, not the PLNet/HAWP work from which the loss is apparently borrowed. This is both a mis-citation and a reproducibility gap: the training objective is not fully specified. The authors should provide the exact loss definitions or at least cite the correct original sources (e.g., HAWP [10] or PLNet, if it is a separate publication) and state any modifications.
[§II-A, Eq. (1)] The decomposition p(y|I) = Σ pJ(yJ|I,yL⁰)·pL(yL|I,yJ⁰) is written as a sum over S(y), but the relationship between the conditional distributions and the set S(y) is only sketched. It is not clear whether S(y) is a set of proposals, a set of latent variables, or a summation over all possible configurations. This is not fatal to the architecture, but it limits the formal precision of the claimed "bidirectional interaction" and should be clarified with a proper generative or discriminative derivation.

minor comments (4)

[§III-A] The endpoint mismatch rate is defined only in words as "the percentage of predicted endpoints without a detected junction within 15 pixels." It should specify the distance metric (Euclidean?), whether the junction is matched to the nearest endpoint, and whether the ground-truth junction set or the predicted set is used.
[§II-C] The sentence "Following HAWP [2]" should cite [10] or [11]; [2] in the reference list is AirSLAM, not HAWP. The same citation issue appears in the loss paragraph, where [2] is again used incorrectly.
[§II-C, Eq. (10)] The text says "all components in our solution are trained end-to-end" but earlier states that the SuperPoint module is fixed. This should be reworded to "all components except SuperPoint" to avoid a direct contradiction.
[Table II] The table formatting is ambiguous: the checkmarks for the PL and LP rows are hard to parse, and the dense-attention row is not clearly indicated. Adding explicit column headers and row separators would improve readability.

Circularity Check

0 steps flagged

No significant circularity; comparisons are external, and the prompt formulation is a conditioning decomposition rather than a self-fulfilling derivation.

full rationale

I walked the claimed derivation chain. Eq. (1) is a stated joint factorization over junction and line predictions conditioned on spatial prompts; it is not a derivation that presupposes the headline result, and the subsequent architecture (PLP-Encoder, CGL-Decoder) is implemented with independently defined losses and external benchmarks. The sAP metric in Eq. (11) uses a squared-endpoint-error criterion that appears inconsistent with the cited 'following [9]' threshold semantics; however, a definitional or reporting inconsistency in the evaluation metric is a correctness/reproducibility risk, not circularity: the model outputs are not fitted to the reported table values, and the comparison numbers for prior methods are not generated from Co-PLNet's own parameters. The paper explicitly omits definitions of L_line, L_junc, and L_aux and defers to Ref. [2] ('Due to space limitations, the definitions ... are not presented here, detailed formulations can be found in [2]'). This is a missing-support/self-citation concern, and Ref. [2] may share authors with the present paper, but it is not load-bearing for the central claim because the contributions (prompts and cross-attention) are not defined in terms of the sAP result, and the comparison includes external baselines (L-CNN, LETR, F-Clip, ELSD, HAWPv2). No step reduces a prediction to its own input by construction. No circularity is therefore established.

Axiom & Free-Parameter Ledger

10 free parameters · 6 axioms · 0 invented entities

The method is a deep-learning architecture; the central claim rests on benchmark metrics, pretrained components, and hand-set hyperparameters. No new physical entities are introduced. The most fragile inherited assumptions are the missing loss definitions and the metric definition.

free parameters (10)

junction confidence threshold = 0.008
Used for junction parsing in PLP-Encoder; inherited from PLNet (Section III-B).
proposal range = 10 pixels
Line proposal search range in HAWP-style parsing (Section III-B).
top proposals retained = 1000
Top-1000 line proposals kept for LOI verification (Section III-B).
prompt channels = 16
Channel width of junction and line prompt encoders (Section III-B).
sparse attention projected channels = 32
Channel dimension after 1x1 convolution in cross-attention (Section III-B).
attention heads = 4
Number of heads in sparse multi-head cross-attention (Section III-B).
attention window size = 8
Spatial window size; selected because it maximized test sAP in Table III, i.e., tuned on test data.
learning rate schedule = 4e-4 for 35 epochs, then 4e-5 for 5 epochs
Optimization schedule (Section III-B).
batch size = 6
Training batch size (Section III-B).
endpoint-junction association radius = not specified
Used to snap endpoints to junctions after decoding; the paper says 'within a certain threshold' without giving the value.

axioms (6)

domain assumption Wireframe and YorkUrban ground-truth labels and the sAP matching rule (Eq. 11) are valid and comparable across methods.
All accuracy claims are computed from these labels and metric; the metric expression is internally inconsistent with standard sAP.
domain assumption The HAWP/HAFM parameterization (d, θ, θ1, θ2, r) is a sufficient representation of line segments for this task.
Line decoding Eq. (4) and line losses rely on it (Sections II-B, III-B).
domain assumption The pretrained SuperPoint weights provide a fixed backbone that transfers to Wireframe/YorkUrban.
SuperPoint is frozen; no fine-tuning is described (Section III-B).
domain assumption The line, junction, and auxiliary losses are exactly those of PLNet/HAWP and are correctly implemented.
The paper does not define these losses; it points to [2], which is AirSLAM, not a wireframe-parsing loss paper.
domain assumption Sparse windowed cross-attention with window size 8 captures sufficient non-local context.
No theoretical guarantee; only empirical ablations (Table III).
domain assumption Gated residual fusion can suppress noise without destabilizing training.
Architectural choice; no evidence provided beyond final performance.

pith-pipeline@v1.3.0-alltime-deepseek · 6731 in / 16357 out tokens · 151929 ms · 2026-08-03T08:01:47.153624+00:00 · methodology

0 comments

read the original abstract

Wireframe parsing aims to recover line segments and their junctions to form a structured geometric representation useful for downstream tasks such as Simultaneous Localization and Mapping (SLAM). Existing methods predict lines and junctions separately and reconcile them post-hoc, causing mismatches and reduced robustness. We present Co-PLNet, a point-line collaborative framework that exchanges spatial cues between the two tasks, where early detections are converted into spatial prompts via a Point-Line Prompt Encoder (PLP-Encoder), which encodes geometric attributes into compact and spatially aligned maps. A Cross-Guidance Line Decoder (CGL-Decoder) then refines predictions with sparse attention conditioned on complementary prompts, enforcing point-line consistency and efficiency. Experiments on Wireframe and YorkUrban show consistent improvements in accuracy and robustness, together with favorable real-time efficiency, demonstrating our effectiveness for structured geometry perception. Our code is available at https://github.com/GalacticHogrider/Co-PLNet.

Figures

Figures reproduced from arXiv: 2601.18252 by Chao Wang, Cheng Dai, Hao Qin, Jinglei Feng, Xuanying Li, Yuqi Ouyang, Yuxiang Luo.

**Figure 2.** Figure 2: For all the channel settings, we refer to Section III-B. ()() [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 2.** Figure 2: Overview of the Co-PLNet framework. The PLP-Encoder generates spatial prompts from junction and line predictions, which are refined by the [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Figure 3: Visualization of parsing results. The first and second rows show images from the Wireframe and YorkUrban datasets, respectively, with predicted [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

discussion (0)

Reference graph

Works this paper leans on

18 extracted references · 1 linked inside Pith

[1]

Airvo: An illumination- robust point-line visual odometry,

K. Xu, Y . Hao, S. Yuan, C. Wang, and L. Xie, “Airvo: An illumination- robust point-line visual odometry,” in2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2023, pp. 3429– 3436

2023
[2]

Airslam: An efficient and illumination-robust point-line visual slam system,

——, “Airslam: An efficient and illumination-robust point-line visual slam system,”IEEE Transactions on Robotics, 2025

2025
[3]

Pluckernet: Learn to register 3d line reconstructions,

L. Liu, H. Li, H. Yao, and R. Zha, “Pluckernet: Learn to register 3d line reconstructions,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 1842–1852

2021
[4]

Learning active basis model for object detection and recognition,

Y . N. Wu, Z. Si, H. Gong, and S.-C. Zhu, “Learning active basis model for object detection and recognition,”International journal of computer vision, vol. 90, no. 2, pp. 198–235, 2010

2010
[5]

Image partitioning into convex polygons,

L. Duan and F. Lafarge, “Image partitioning into convex polygons,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2015, pp. 3119–3127

2015
[6]

Instance segmentation with mask- supervised polygonal boundary transformers,

J. Lazarow, W. Xu, and Z. Tu, “Instance segmentation with mask- supervised polygonal boundary transformers,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 4382–4391

2022
[7]

Fully convolutional line parsing,

X. Dai, H. Gong, S. Wu, X. Yuan, and Y . Ma, “Fully convolutional line parsing,”Neurocomputing, vol. 506, pp. 1–11, 2022

2022
[8]

Line segment detection using transformers without edges,

Y . Xu, W. Xu, D. Cheung, and Z. Tu, “Line segment detection using transformers without edges,” inProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, 2021, pp. 4257–4266

2021
[9]

End-to-end wireframe parsing,

Y . Zhou, H. Qi, and Y . Ma, “End-to-end wireframe parsing,” inProceed- ings of the IEEE/CVF International Conference on Computer Vision, 2019, pp. 962–971

2019
[10]

Holistically-attracted wireframe parsing,

N. Xue, T. Wu, S. Bai, F. Wang, G. Xia, L. Zhang, and P. H. S. Torr, “Holistically-attracted wireframe parsing,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 2788–2797

2020
[11]

Holistically-attracted wireframe parsing: From supervised to self-supervised learning,

——, “Holistically-attracted wireframe parsing: From supervised to self-supervised learning,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 12, pp. 14 727–14 744, 2023

2023
[12]

U-net: Convolutional networks for biomedical image segmentation,

O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” inInternational Conference on Medical image computing and computer-assisted intervention. Springer, 2015, pp. 234–241

2015
[13]

Superpoint: Self- supervised interest point detection and description,

D. DeTone, T. Malisiewicz, and A. Rabinovich, “Superpoint: Self- supervised interest point detection and description,” inProceedings of the IEEE conference on computer vision and pattern recognition workshops, 2018, pp. 224–236

2018
[14]

Learning to parse wireframes in images of man-made environments,

K. Huang, Y . Wang, Z. Zhou, T. Ding, S. Gao, and Y . Ma, “Learning to parse wireframes in images of man-made environments,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018, pp. 626–635

2018
[15]

Efficient edge-based methods for estimating manhattan frames in urban imagery,

P. Denis, J. H. Elder, and F. J. Estrada, “Efficient edge-based methods for estimating manhattan frames in urban imagery,” inEuropean conference on computer vision. Springer, 2008, pp. 197–210

2008
[16]

A method for stochastic optimization,

D. P. Kingma and J. Ba, “A method for stochastic optimization,”arXiv preprint arXiv:1412.6980, vol. 1412, no. 6, 2014

Pith/arXiv arXiv 2014
[17]

Learning attraction field representation for robust line segment detection,

N. Xue, S. Bai, F. Wang, G. Xia, T. Wu, and L. Zhang, “Learning attraction field representation for robust line segment detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 1595–1603

2019
[18]

Elsd: Efficient line segment detector and descriptor,

H. Zhang, Y . Luo, F. Qin, Y . He, and X. Liu, “Elsd: Efficient line segment detector and descriptor,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 2969–2978

2021

[1] [1]

Airvo: An illumination- robust point-line visual odometry,

K. Xu, Y . Hao, S. Yuan, C. Wang, and L. Xie, “Airvo: An illumination- robust point-line visual odometry,” in2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2023, pp. 3429– 3436

2023

[2] [2]

Airslam: An efficient and illumination-robust point-line visual slam system,

——, “Airslam: An efficient and illumination-robust point-line visual slam system,”IEEE Transactions on Robotics, 2025

2025

[3] [3]

Pluckernet: Learn to register 3d line reconstructions,

L. Liu, H. Li, H. Yao, and R. Zha, “Pluckernet: Learn to register 3d line reconstructions,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 1842–1852

2021

[4] [4]

Learning active basis model for object detection and recognition,

Y . N. Wu, Z. Si, H. Gong, and S.-C. Zhu, “Learning active basis model for object detection and recognition,”International journal of computer vision, vol. 90, no. 2, pp. 198–235, 2010

2010

[5] [5]

Image partitioning into convex polygons,

L. Duan and F. Lafarge, “Image partitioning into convex polygons,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2015, pp. 3119–3127

2015

[6] [6]

Instance segmentation with mask- supervised polygonal boundary transformers,

J. Lazarow, W. Xu, and Z. Tu, “Instance segmentation with mask- supervised polygonal boundary transformers,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 4382–4391

2022

[7] [7]

Fully convolutional line parsing,

X. Dai, H. Gong, S. Wu, X. Yuan, and Y . Ma, “Fully convolutional line parsing,”Neurocomputing, vol. 506, pp. 1–11, 2022

2022

[8] [8]

Line segment detection using transformers without edges,

Y . Xu, W. Xu, D. Cheung, and Z. Tu, “Line segment detection using transformers without edges,” inProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, 2021, pp. 4257–4266

2021

[9] [9]

End-to-end wireframe parsing,

Y . Zhou, H. Qi, and Y . Ma, “End-to-end wireframe parsing,” inProceed- ings of the IEEE/CVF International Conference on Computer Vision, 2019, pp. 962–971

2019

[10] [10]

Holistically-attracted wireframe parsing,

N. Xue, T. Wu, S. Bai, F. Wang, G. Xia, L. Zhang, and P. H. S. Torr, “Holistically-attracted wireframe parsing,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 2788–2797

2020

[11] [11]

Holistically-attracted wireframe parsing: From supervised to self-supervised learning,

——, “Holistically-attracted wireframe parsing: From supervised to self-supervised learning,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 12, pp. 14 727–14 744, 2023

2023

[12] [12]

U-net: Convolutional networks for biomedical image segmentation,

O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” inInternational Conference on Medical image computing and computer-assisted intervention. Springer, 2015, pp. 234–241

2015

[13] [13]

Superpoint: Self- supervised interest point detection and description,

D. DeTone, T. Malisiewicz, and A. Rabinovich, “Superpoint: Self- supervised interest point detection and description,” inProceedings of the IEEE conference on computer vision and pattern recognition workshops, 2018, pp. 224–236

2018

[14] [14]

Learning to parse wireframes in images of man-made environments,

K. Huang, Y . Wang, Z. Zhou, T. Ding, S. Gao, and Y . Ma, “Learning to parse wireframes in images of man-made environments,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018, pp. 626–635

2018

[15] [15]

Efficient edge-based methods for estimating manhattan frames in urban imagery,

P. Denis, J. H. Elder, and F. J. Estrada, “Efficient edge-based methods for estimating manhattan frames in urban imagery,” inEuropean conference on computer vision. Springer, 2008, pp. 197–210

2008

[16] [16]

A method for stochastic optimization,

D. P. Kingma and J. Ba, “A method for stochastic optimization,”arXiv preprint arXiv:1412.6980, vol. 1412, no. 6, 2014

Pith/arXiv arXiv 2014

[17] [17]

Learning attraction field representation for robust line segment detection,

N. Xue, S. Bai, F. Wang, G. Xia, T. Wu, and L. Zhang, “Learning attraction field representation for robust line segment detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 1595–1603

2019

[18] [18]

Elsd: Efficient line segment detector and descriptor,

H. Zhang, Y . Luo, F. Qin, Y . He, and X. Liu, “Elsd: Efficient line segment detector and descriptor,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 2969–2978

2021