Pith. sign in

REVIEW 4 major objections 5 minor 18 references

Advancing Utility Pole and Sign Detection Through Deep Learning

T0 review · 4 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read A single street photo can localize a wooden utility pole and estimate its lean angle to within about a degree.

desk verdict The dataset is a real contribution, but the unaddressed train/test leakage from multiple images per pole makes the headline numbers provisional; the paper needs major revision before its accuracy claims can be trusted. read the letter →

arxiv 2608.04061 v1 pith:NQQOA4GB submitted 2026-08-04 cs.CV

Carl Dickinson , Gaetano Di Caterina This is my paper · ORCID
classification cs.CV
keywords utilitypoledetectionleanangleestimationtransformerinstancesegmentationoverheadlineinspectionstreet-viewimageryinfrastructuremonitoringobjectbenchmark
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

Wooden utility poles carry much of the distribution grid, but inspecting them still relies on foot patrols and aerial surveys. This paper argues that a single transformer-based detection model, trained on street-level imagery, can simultaneously find poles, recognize attached warning signs, and estimate how far each pole leans. On a held-out test set the model reports 90.43% mean average precision for poles, 88.26% for signs, and a lean-angle mean absolute error of 1.01 degrees, with 98% of estimates within 5 degrees. The authors also release their 4,570-image annotated corpus as a public benchmark, so the value of the claim is that a cheap, repeatable ground-level screen could triage which poles need a physical visit.

What carries the argument

The load-bearing object is DETR, a transformer-based detector that predicts a fixed set of object queries with a bipartite matching loss, eliminating hand-crafted anchors and non-maximum suppression. The paper extends its decoder with a segmentation head that upsamples attention maps into per-instance masks. For each mask, a line-fit routine computes the dominant orientation, and the pole's lean angle is defined by $\theta = \arctan(\mathrm{rise}/\mathrm{run})$ in degrees. This shared representation is what lets detection and angle estimation come from one end-to-end trained model.

What would settle it

Physically measure the lean of a sample of poles (or obtain LiDAR-derived tilt), photograph each pole from several street-level headings, and compare the model's mask-derived angles with the physical measurements. If the mean absolute error is substantially larger than 1.01 degrees, or rises with oblique camera headings, the central lean-angle claim is an artifact of mask-to-mask agreement rather than true tilt.

Watch

Extended reading notes

Core claim

The central claim is that the mass of a pole in the image, formalized as a per-instance segmentation mask, carries enough orientation information to serve as a lean estimate, so no extra sensor or multi-stage heuristic is needed. The paper's implementation attaches a segmentation head to a detection transformer, fits a line to each predicted pole mask, and reports a 1.01 degree mean absolute error on the 1,367 of 1,433 test-set poles for which an angle could be produced. It further reports that 70.5% of predictions are within 1 degree and 98% within 5 degrees, and that on a filtered subset with pole heights comparable to earlier studies, its within-1-degree accuracy exceeds three prior UAV-based pole-angle methods by 11.8 to 22.3 percentage points. The paper also compares with a more recent segmentation detector under the same protocol; that detector achieves a lower angle error (0.645 degrees) and higher pole F1 at IoU 0.5 on its own masks, a result the authors read as a coverage-versus-sharpness trade-off between the two models.

Load-bearing premise

The lean-angle accuracy tallies rely on treating the angle of the pole's outline in the photo, produced by a line fit over the annotated or predicted mask, as the pole's true tilt; perspective distortion and annotation noise could make the 1.01-degree error a measure of label agreement rather than physical lean.

Editorial extensions

If this is right

  • A utility could screen its overhead-line network from existing street-level imagery, flagging poles whose estimated lean exceeds a threshold for a follow-up visit.
  • Warning-sign classification runs in the same forward pass, so asset inventories can be updated as a by-product of pole surveys.
  • The public OHL-UK corpus gives later detectors a fixed benchmark on long, thin, visually ambiguous wooden poles, making cross-model results directly comparable.
  • The paper's comparison shows the same mask-to-angle pipeline transfers to another segmentation detector, which produced a lower angle error on its own masks; this makes the angle result look like a property of good masks rather than of one architecture.

Reading between the lines

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

  • Because the dataset was captured at four compass headings per location, the effect of viewpoint on angle accuracy can be tested directly: if errors grow for oblique headings, the 2D-mask angle is partly a perspective artifact, and multi-view fusion would be the natural correction.
  • The ground-truth labels themselves are mask-based, so the reported accuracy should be read as an upper bound on physical-lean accuracy; a LiDAR or manual plumbline validation would likely reveal a bias term that this evaluation cannot see.
  • The same detection-plus-segmentation-plus-line-fit recipe should transfer to other long, thin infrastructure objects such as streetlights, masts, and fence lines, where lean or tilt is a useful health indicator.
  • Where both DETR and the competing segmentation model produced masks, the paper's own numbers suggest fusing their angles could improve overall triage accuracy.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper presents a detection, segmentation, and lean-angle-estimation framework for wooden utility poles and attached warning signs in Google Street View imagery. The authors introduce OHL-UK, a 4,570-image dataset with 6,773 pole and 1,805 sign instances annotated with boxes, masks, and mask-derived lean angles. A fine-tuned DETR with a segmentation head is trained and compared with RetinaNet, Faster R-CNN, YOLOv3-Tiny, and later YOLOv8 and DINO-DETR. The abstract claims 90.43% pole mAP, 88.26% sign mAP, and a 1.01 degree mean absolute lean-angle error. The main contributions are stated as the dataset, a DETR-based detector, and an integrated lean-angle pipeline.

Significance. The dataset release and the application are valuable for infrastructure inspection. The paper is also careful to compare with modern detectors in Section 4, and it candidly notes in Section 4.3 that the lean-angle ground truth is mask-derived rather than absolute tilt. However, the headline claims are not currently supported: the abstract's mAP values are not reported in the tables, the state-of-the-art claim is contradicted by the paper's own YOLOv8 comparison, and the image-level split may leak the same physical poles into train and test sets. If the model is re-evaluated on a location-grouped split and the claims are recalibrated, the work could be a solid applied contribution.

major comments (4)
  1. [3.3 and Table 1] Section 3.3 states that the test set comprises 920 images with 6,773 wooden utility poles and 1,805 warning signs, but Table 1 reports only 1,490 poles and 552 signs in the test split; the TP+FN counts at IoU 0.0 in Table 3 (1,433+57 and 542+10) confirm that the test set contains 1,490 poles and 552 signs. This is an internal inconsistency in the central evaluation description and must be fixed.
  2. [Abstract and Section 4.2] The abstract and introduction claim state-of-the-art detection and angle estimation, but Section 4.2's Table 6 shows YOLOv8 achieving F1 0.93 at IoU 0.5 on poles versus DETR's 0.86, and Table 8 shows YOLOv8 with +44 TP, -119 FP, and -109 FN relative to DETR; Section 4.3 further reports YOLOv8-seg MAE 0.645 degrees versus DETR-seg 1.01 degrees. The paper cannot claim state-of-the-art or 'outperforms' without reconciling these numbers, and the abstract's 90.43/88.26 mAP values do not appear in any table.
  3. [2.1 and 3.1] The data collection captures four images per geographic coordinate at headings 0, 90, 180, and 270 degrees, so the same physical pole appears in multiple images. Section 3.1 describes only stratified image-level splits and gives no grouping by coordinate or pole identity. This permits near-duplicate images of the same pole to appear in both training and test splits, potentially inflating detection and angle metrics. The authors should quantify leakage, for example by reporting coordinate overlap between splits, and report results on a coordinate-grouped split.
  4. [2.1 and 4.3] Lean-angle ground truth is obtained by applying fitLine to the manually annotated mask, and the predicted angle is obtained by applying the same fitLine to the predicted mask. Because the network is trained to reproduce those masks, the reported MAE measures mask-to-mask agreement rather than absolute physical lean, as the authors concede in Section 4.3. The claims should be reworded accordingly, or validated against independently measured pole lean.
minor comments (5)
  1. [3.1] Section 3.1 reports a best mean mAP of 90.46% from the ablation, while the abstract reports 90.43%; clarify which number corresponds to the final test configuration and which folds.
  2. [3.2 and Table 2] Table 2 reports 'DETR 90%' but the text in Section 3.2 says 'reaching 90% mAP'; specify the reported precision, for example AP@0.5 versus COCO-style mAP, consistently across all tables.
  3. [4.1] In Section 4.1 the phrase 'essentially tied' is used for AP@0.5 values of 90.0, 89.7, and 89.3; report variance across folds since the differences are within normal noise.
  4. [3.4] Section 3.4's filtered subset of 642 poles is selected by height to match previous studies; specify the filter criterion and whether this subset was chosen before or after seeing the angle errors.
  5. [Abstract] The abstract says 'model accurately estimates lean for 1,367 out of 1,433 test-set poles,' but this count is defined at IoU 0.0 matching; state the matching criterion in the abstract or use a more standard threshold.

Circularity Check

1 steps flagged · score 6.0 of 10

Lean-angle metric is self-referential: fitLine on the annotation defines the ground truth and fitLine on the predicted mask defines the prediction; detection claims remain independent.

  1. self definitional [Sections 2.1, 2.2 (Eq. 1), 3.4, 4.3]
    "Each pole annotation is additionally labelled with a lean angle in degrees, computed by fitting a line to the segmentation mask using OpenCV’s fitLine() method. ... Estimated pole masks are post-processed using OpenCV’s fitLine() function to obtain the dominant orientation vector. From this, the lean angle θ is computed as: θ=arctan(rise/run) ... GT angles come from mask-derived fitLine estimates, so errors are mask-to-mask, not absolute tilt."

    The ground-truth lean angle is not an independent physical measurement; it is a deterministic function of the annotation mask (fitLine on the polygon). The predicted lean angle is the same deterministic function applied to the predicted mask. Since the model is trained to reproduce annotation masks, the reported 1.01° MAE is a re-expression of segmentation fidelity, not an independent estimate of physical pole lean. The paper explicitly concedes this in Section 4.3: 'errors are mask-to-mask, not absolute tilt.' Thus the lean-angle 'prediction' reduces by construction to mask agreement with the annotations.

full rationale

The detection and sign-classification results (pole mAP 90.43%, sign mAP 88.26%) are evaluated against independently annotated bounding boxes, so those claims are not circular. The single reduction-by-construction step is the lean-angle evaluation: Eq. (1) converts a 2D mask to an angle, and the same OpenCV fitLine() operation is applied both to annotation masks to form ground-truth angles and to predicted masks to form predicted angles. Because the segmentation head is trained to reproduce the annotation masks, the angle error is essentially a mask-to-mask agreement metric. The paper admits this in Section 4.3, but the abstract nonetheless presents the 1.01° MAE as an accurate estimate of physical pole lean. The train/test split concern (four headings per coordinate, no grouping by pole identity) is a potentially serious data-leakage issue, but it is not a circularity and therefore is not scored here. The self-citation to the released dataset [5] is a normal data availability citation and is not load-bearing. Overall, the central detection claims have independent content, but one headline quantity — lean-angle estimation — is partially circular by construction, warranting a score of 6.

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

The central detection results rest on standard supervised learning: the model weights are fitted to a manually annotated dataset, which is legitimate. The most ledger-relevant choices are the hyperparameters tuned by ablation and the filtered 642-pole subset used for prior-work comparison. The lean-angle claim rests on a domain assumption that a 2D mask fitLine angle is a valid lean ground truth; this is the weakest link. No new physical entities are introduced.

free parameters (3)
  • DETR learning rate = 1e-5
    Selected by an ablation study on the validation split (Section 3.1). Not a theoretical constant; the reported mAP depends on this choice and could differ with other tuning.
  • DETR weight decay = 1e-4
    Selected by the same ablation (Section 3.1).
  • Filtered evaluation subset height window = mean height 346.9 px, 642 poles
    Used in Section 3.4 for the favorable comparison with prior work; chosen to match height distribution of previous studies, which is a post hoc selection that inflates accuracy (92.8% within 1 deg vs 70.5% on the full test set).
assumptions (3)
  • domain assumption ImageNet-pretrained ResNet-50 weights transfer to street-view utility pole imagery
    Stated in Section 2.3 as initialization for DETR; no evidence is provided that pretraining on ImageNet is optimal for this domain, and it is a background assumption the results rely on.
  • domain assumption The angle of the line fit to the 2D mask equals the physical lean of the pole
    Introduced in Section 2.1 and Equation (1); the entire lean-angle evaluation treats mask-derived fitLine angles as ground truth, and the paper later concedes the metric is mask-to-mask (Section 4.3).
  • domain assumption Google Street View imagery at four compass headings provides a representative sample of UK wooden poles without crossarms
    Dataset construction in Section 2.1 assumes these images, filtered manually, are a fair sample for training and benchmarking; the paper itself lists domain generalization as a limitation (Section 5).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Advancing Utility Pole and Sign Detection Through Deep Learning." pith.science (2026). https://pith.science/paper/NQQOA4GB

@misc{pith2026260804061,
  author       = {Pith},
  title        = {Pith review of: Advancing Utility Pole and Sign Detection Through Deep Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NQQOA4GB}},
  note         = {Machine review of arXiv:2608.04061}
}
read the original abstract

Utility poles are an essential part of the infrastructure used to support power distribution systems and other critical public services. Their regular inspection is crucial to ensure the stability and safety of the electrical grid. A deep learning framework is presented for the automated detection, segmentation and lean angle estimation of wooden utility poles, and classification of attached electrical warning signs, using ground-level imagery. The system is trained on a custom dataset of 4,570 annotated images extracted from Google Street View, featuring challenging real-world scenes with visually ambiguous wooden poles lacking distinctive features. The proposed model is based on the Detection Transformer (DETR), suitably modified and trained on the custom dataset. The model outperforms standard object detectors (RetinaNet, Faster R-CNN, YOLOv3-Tiny), achieving a mean average precision of 90.43% for pole detection and 88.26% for sign detection. Extending this model with a segmentation head enables per-instance mask generation, which is then used to estimate pole lean angle. The model accurately estimates lean for 1,367 out of 1,433 test-set poles, with a mean absolute error of 1.01 degrees. Moreover, the custom dataset created in this work is also made publicly available to be used as a benchmark.

Figures

Figures reproduced from arXiv: 2608.04061 by the authors.

Figure 1
Figure 1. Examples of a wooden utility pole and an electrical warning sign from the OHL [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. DETR architecture: a convolutional backbone extracts features from the input [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Segmentation head added to DETR, enabling binary mask prediction for each [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Example DETR detection output showing confident localisation of a wooden utility [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Examples of segmented utility poles with predicted confidence scores and esti [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

18 extracted references · 10 canonical work pages

  1. [1]

    ac.uk/user-guide/

    Introduction – ARCHIE-WeSt documentation.https://docs.hpc.strath. ac.uk/user-guide/. Accessed 2025

    work page 2025

  2. [2]

    Md Morshedul Alam, Zanbo Zhu, Berna Eren Tokgoz, Jing Zhang, and Seokyon Hwang. Automatic assessment and prediction of the resilience of utility poles using unmanned aerial vehicles and computer vision techniques.International Journal of Disaster Risk Science, 11(1):119–132, 2020.10.1007/s13753-020-00254-1. URLhttps://doi.org/10.1007/s13753-020-00254-1

    work page doi:10.1007/s13753-020-00254-1 2020

  3. [3]

    End-to-end object detection with transformers

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kir- illov, and Sergey Zagoruyko. End-to-end object detection with transformers. In An- drea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors,Computer Vision – ECCV 2020, volume 12346 ofLecture Notes in Computer Science, pages 213–229. Springer, 2020.10.1007/9...

    doi:10.1007/978-3-030-58452-8_13 2020

  4. [4]

    Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V

    Ekin D. Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V . Le. Autoaugment: Learning augmentation strategies from data. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 113–123, 2019.10.1109/CVPR.2019.00020

    arXiv 2019

  5. [5]

    OHL-UK: Wooden utility pole and electrical sign corpus with trained detection and segmentation models, 2025

    Carl Dickinson. OHL-UK: Wooden utility pole and electrical sign corpus with trained detection and segmentation models, 2025. URLhttps://doi.org/10.15129/ df7cc895-1091-4cd6-bf2c-7b5ebd203e55. Dataset openly available under CC BY 4.0 licence

    work page 2025

  6. [6]

    The VGG image annotator (VIA).arXiv preprint arXiv:1904.10699, 2019

    Abhishek Dutta and Andrew Zisserman. The VGG image annotator (VIA).arXiv preprint arXiv:1904.10699, 2019. URLhttps://arxiv.org/abs/1904. 10699

    arXiv 1904

  7. [7]

    Review and latest trends in mobile robots used on power transmission lines.International Journal of Advanced Robotic Systems, 10: 1–14, 2013.10.5772/56791

    Rogério Gonçalves and João Carvalho. Review and latest trends in mobile robots used on power transmission lines.International Journal of Advanced Robotic Systems, 10: 1–14, 2013.10.5772/56791. URLhttps://doi.org/10.5772/56791

    work page doi:10.5772/56791 2013

  8. [8]

    Jaeyoon Kim, Mirsalar Kamari, Seulbi Lee, and Youngjib Ham. Large-scale visual data–driven probabilistic risk assessment of utility poles regarding the vulnerability of power distribution infrastructure systems.Journal of Construction Engineering and Management, 147(10):04021121, 2021.10.1061/(ASCE)CO.1943-7862. 0002153

    work page doi:10.1061/(asce)co.1943-7862 2021

Show all 18 references
  1. [9]

    Deep inspection: An electrical distribution pole parts study VIA deep 12DICKINSON & DI CA TERINA: UTILITY POLE AND W ARNING SIGN DETECTION neural networks

    Liangchen Liu, Teng Zhang, Kun Zhao, Arnold Wiliem, Kieren Astin-Walmsley, and Brian Lovell. Deep inspection: An electrical distribution pole parts study VIA deep 12DICKINSON & DI CA TERINA: UTILITY POLE AND W ARNING SIGN DETECTION neural networks. In2019 IEEE International Co...

    2019

  2. [10]

    Learning data augmentation with online bilevel optimization for image classification

    Saypraseuth Mounsaveng, Issam Laradji, Ismail Ben Ayed, David Vazquez, and Marco Pedersoli. Learning data augmentation with online bilevel optimization for image classification. In2021 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1690–1699, 2021.10.1...

    2021

  3. [11]

    Adithya, Tanima Dutta, and P

    Hrishikesh Sharma, V . Adithya, Tanima Dutta, and P. Balamuralidhar. Image analysis- based automatic utility pole detection for remote surveillance. In2015 International Conference on Digital Image Computing: Techniques and Applications (DICTA), pages 1–7. IEEE, 2015.10.1109/D...

    2015

  4. [12]

    Future smart: Consultation report

    UK Power Networks. Future smart: Consultation report. Techni- cal report, UK Power Networks, 2017. URLhttps://media. umbraco.io/uk-power-networks/2etfstsb/futuresmart_ consultationreport_.pdf

    2017

  5. [13]

    google_streetview: Google street view image api command-line tool and python module (v1.2.3).https://rrwen.github.io/google_ streetview/

    Richard Wen. google_streetview: Google street view image api command-line tool and python module (v1.2.3).https://rrwen.github.io/google_ streetview/. Accessed 2025

    2025

  6. [14]

    C. C. Whitworth, A. W. G. Duller, D. I. Jones, and G. K. Earp. Aerial video inspection of overhead power lines.Power Engineering Journal, 15(1):25–32, 2001.10.1049/ pe:20010103. URLhttps://digital-library.theiet.org/doi/10. 1049/pe%3A20010103

    2001

  7. [15]

    What is YOLOv8: An in-depth exploration of the internal features of the next-generation object detector, 2024

    Muhammad Yaseen. What is YOLOv8: An in-depth exploration of the internal features of the next-generation object detector, 2024. URLhttps://arxiv.org/abs/ 2408.15857

    2024 arXiv

  8. [16]

    Ni, and Heung-Yeung Shum

    Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M. Ni, and Heung-Yeung Shum. DINO: DETR with improved denoising anchor boxes for end-to- end object detection, 2022. URLhttps://arxiv.org/abs/2203.03605

    2022 arXiv

  9. [17]

    Using deep learning to identify utility poles with crossarms and estimate their locations from google street view images.Sensors, 18(8):2484, 2018.10.3390/ s18082484

    Weixing Zhang, Chandi Witharana, Weidong Li, Chuanrong Zhang, Xiaojiang Li, and Jason Parent. Using deep learning to identify utility poles with crossarms and estimate their locations from google street view images.Sensors, 18(8):2484, 2018.10.3390/ s18082484. URLhttps://www.m...

    2018

  10. [18]

    Automatic utility pole inclination angle measurement using unmanned aerial vehicle and deep learning

    Zanbo Zhu, Jing Zhang, Md Morshedul Alam, Berna Eren Tokgoz, and Seokyon Hwang. Automatic utility pole inclination angle measurement using unmanned aerial vehicle and deep learning. InIISE Annual Conference and Expo 2019, 2019

    2019

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.