Pith. sign in

REVIEW 3 major objections 4 minor 33 references

A full-resolution recurrent network trained only on synthetic noisy rail rasters cleans classified lidar into vectorizable tracks and centerlines with almost no hand editing.

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 · grok-4.5

2026-07-10 20:26 UTC pith:UJJX77HE

load-bearing objection Solid engineering pipeline for rail centerlines from classified MLS; FRPDF is a real full-resolution recurrent variant, but real-world evidence is only one qualitative example and synthetic F1, so operational claims stay provisional. the 3 major comments →

arxiv 2607.06829 v1 pith:UJJX77HE submitted 2026-07-07 cs.CV

Rail Track Extraction from Rasterized Classified Point Clouds Using a Full-Resolution, Fully Convolutional Recurrent Neural Network

Alexander Gribov , Jie Chang This is my paper
classification cs.CV
keywords rail track extractionmobile laser scanningsynthetic training datafull-resolution CNNrecurrent dilated fusionvectorizationcenterline reconstructionpoint cloud rasterization
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.

Rail operators need accurate 3-D rail-top and track centerlines for safety systems and asset maps, but classified lidar still contains gaps, density variation and false positives that break ordinary vectorization. This paper shows that a lightweight fully convolutional recurrent network kept at full spatial resolution, trained exclusively on procedurally generated noisy rail rasters, can suppress that noise and restore continuous parallel rails. After morphological closing, classical vectorization, geometric smoothing and Dynamic Time Warping pairing, the cleaned raster yields usable 3-D polylines. Because the training data never touch proprietary rail surveys, the method removes the usual labeling bottleneck while remaining practical for large corridors.

Core claim

The Full-Resolution Progressive Dilated Fusion (FRPDF) architecture, run recurrently for two iterations and trained only on synthetic rasters that deliberately include shifts, gaps, thickness jitter, pixel flips and random blobs, produces binary rail maps clean enough for reliable vectorization and subsequent 3-D centerline reconstruction from real classified MLS point clouds.

What carries the argument

FRPDF: a full-resolution, fully convolutional recurrent stack of progressive dilated fusion blocks (dilations 1,3,7,...,127) that aggregates multiscale context without down-sampling, then feeds its own output back for a second refinement pass.

Load-bearing premise

The synthetic noise model plus a handful of gauge-and-thickness pairs is assumed to cover the real density drops, misclassifications, occlusions and geometry that appear in mobile lidar of actual railways.

What would settle it

Run the same pretrained models on a held-out real multitrack corridor that contains dense vegetation, switch complexes or extreme density fall-off; if the cleaned raster still leaves large unbridgeable gaps or invents false rails that survive morphological closing and DTW pairing, the generalization claim fails.

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

If this is right

  • Once rail points are classified, the remaining extraction pipeline can run end-to-end without further manual editing on ordinary multi-track corridors.
  • New track gauges or rail-head thicknesses can be handled by selecting or lightly scaling an existing synthetic model rather than collecting and labeling fresh field data.
  • The same full-resolution recurrent cleaner can be reused for other thin, continuous linear features whose local geometry is regular enough to be synthesized.
  • Railway mapping workflows can replace proprietary labeled training sets with procedurally generated rasters, cutting both cost and access barriers.

Where Pith is reading between the lines

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

  • The same synthetic-plus-full-resolution pattern should transfer to other infrastructure linework (road markings, pipelines, power lines) whenever the target has strong geometric regularity and the dominant noise can be procedurally mimicked.
  • If turnouts and crossings are added to the synthetic generator, the current post-processing DTW stage may need only modest extension rather than an entirely new architecture.
  • Because the network never sees real labels, any residual domain gap will appear as systematic failure modes (e.g., vegetation blobs that look like short rail stubs) that can be diagnosed by comparing synthetic versus real confusion maps.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes FRPDF, a full-resolution fully convolutional recurrent network with progressive dilated fusion, trained exclusively on procedurally generated synthetic noisy rail rasters, to clean binary rasters of classified MLS rail points. After morphological closing, veinerization-based vectorization, polyline smoothing, z-transfer from lidar, and DTW-based rail pairing, the pipeline produces rail-top centerlines and track centerlines with claimed minimal manual intervention. Synthetic test metrics (precision 0.809, recall 0.946, F1 0.872) and one qualitative multi-track real-data example are presented as validation; the method is implemented in ArcGIS Pro.

Significance. If the claimed generalization from synthetic noise to real MLS holds, the work offers a practical, data-efficient pipeline for automated rail geometry extraction that sidesteps proprietary labeled datasets and downsampling artifacts. Strengths include the explicit full-resolution design motivated by vectorization needs, the detailed synthetic generation process covering gauge/thickness variation and multiple noise types, the recurrent refinement loop, and the end-to-end integration with existing geometric post-processing (smoothing, DTW, extremal-boundary selection). These elements make the approach potentially useful for railway asset mapping workflows even if quantitative real-world error bars remain to be established.

major comments (3)
  1. [§6, Table 1, §10] §6 / Table 1 report only synthetic confusion-matrix metrics on 1024×1024 images; §10 and Fig. 8 provide a single qualitative multi-track example with no measured lateral/vertical RMSE, completeness, or false-positive rates against surveyed ground truth. Prior geometric methods cited in §2 claim 2–3 cm accuracy; without comparable real-data numbers the central claim of “high-quality” operational extraction (Abstract, §11.1) remains unquantified and cannot be assessed against the literature.
  2. [§1, §6] The authors explicitly decline a capacity-matched ACNN baseline (§1, §6) because fair re-engineering was judged out of scope. Given that FRPDF is presented as a novel architecture whose full-resolution recurrent design is the key technical contribution, the absence of any competitive quantitative comparison (even on the synthetic test set) leaves the claimed advantage over existing full-resolution or dilated architectures unsubstantiated.
  3. [§4, §7, §11.2] The load-bearing assumption that the synthetic noise model (rail shifts, missing segments, thickness jitter, pixel flips, random blobs; §4) plus sparse gauge/thickness pairs is representative of real MLS density variation, misclassification, occlusion and complex geometry is stated but not tested. §11.2 itself lists turnouts, crossings and realistic density/occlusion patterns as future work; if those modes differ substantially from the synthetic distribution, the claimed robustness without real labeled training data does not hold.
minor comments (4)
  1. [§5, Fig. 7] Figure 7’s recurrent decision diamond and dual thresholds (1/2 and 3/4) are described only schematically; the precise stopping criterion and whether the number of iterations is fixed or data-dependent should be stated more formally.
  2. [§9.1, Appendix] The appendix derives the integral of squared distance used as DTW penalty, yet the main text (§9.1) never references the appendix equation; a forward pointer would improve readability.
  3. [§3] Several ArcGIS tool names and version numbers appear without corresponding open algorithmic descriptions; readers outside the Esri ecosystem would benefit from brief pseudocode or parameter lists for the interactive rail selection and thinning steps (§3.2).
  4. [throughout] Typographical inconsistencies appear (e.g., “T raining”, “V ectorized”, “F uture W ork”); a final copy-edit pass is needed.

Circularity Check

0 steps flagged

No significant circularity: empirical method paper whose NN is trained and scored independently of its post-processing tools.

full rationale

The paper's load-bearing claim is that FRPDF (full-resolution progressive dilated fusion), trained exclusively on procedurally generated noisy synthetic rail rasters (§4–5), produces cleaned binary rasters that, after morphological closing, veinerization, smoothing and DTW pairing, yield usable rail-top and track centerlines from classified MLS clouds (Abstract, §7–10, §11.1). Training targets are generated by an independent geometric pipeline (ideal centerlines + deliberate distortions + pixel noise); evaluation uses ordinary confusion-matrix metrics on held-out synthetic images (Table 1) plus one qualitative real-data example (Fig. 8). No equation or claimed prediction reduces by construction to a fitted free parameter. Self-citations ([1–3], [17], [29]) point to previously published vectorization/smoothing utilities that are applied as ordinary post-processing steps; they do not supply uniqueness theorems, ansatzes, or uniqueness results that force the central FRPDF claim. The derivation chain is therefore self-contained against its own inputs.

Axiom & Free-Parameter Ledger

5 free parameters · 3 axioms · 1 invented entities

The central claim rests on architectural hyper-parameters chosen by the authors, domain assumptions about rail geometry regularity, and the unproven representativeness of the synthetic noise process. No new physical entities are postulated; the invented entity is the FRPDF module itself.

free parameters (5)
  • feature channel count C = 24
    Fixed at 24 throughout encoder/bridge/decoder; chosen to control compute while preserving full resolution (§5, Fig. 6).
  • recurrent iteration count = 2
    Fixed at two FRPDF passes; authors note further iterations help less (§5, Fig. 7).
  • binary decision thresholds = 0.5 / 0.75
    1/2 for final output, 3/4 for feedback loop (§5, Fig. 7).
  • raster cell size and synthetic gauge/thickness pairs = 4 cm; 1.435 m / 76.2 mm (example)
    Cell size ~4 cm for US 132 RE example; gauge and railhead thickness drawn from standards plus sparse extra pairs for generalization (§4, §7).
  • synthetic noise probabilities and blob parameters
    Randomly sampled pixel-flip rates and blob sizes/numbers used to corrupt training rasters (§4.3); exact distributions not tabulated.
axioms (3)
  • domain assumption Rails are locally continuous, roughly parallel at constant gauge, and of limited curvature variation, so a full-resolution dilated network can recover them from noisy rasters.
    Stated as the regime of applicability in the introduction and conclusion; method is explicitly not intended for abrupt random direction changes.
  • ad hoc to paper Procedurally generated synthetic rasters with the described distortions are a sufficient training distribution for real MLS rail rasters.
    Core training strategy (§4, §7); no real labeled rails are used for training.
  • standard math Standard dilated-convolution and batch-norm building blocks behave as in the cited literature.
    Architecture re-uses Conv2d + BatchNorm2d + ReLU and progressive dilation rates from prior CV work.
invented entities (1)
  • FRPDF (Full-Resolution Progressive Dilated Fusion) layer and its recurrent wrapper no independent evidence
    purpose: Clean noisy rail rasters at full spatial resolution while aggregating multi-scale context without downsampling or skip connections.
    Defined in §5 with block diagrams; independent evidence is limited to synthetic metrics and one qualitative real example.

pith-pipeline@v1.1.0-grok45 · 17879 in / 2963 out tokens · 39524 ms · 2026-07-10T20:26:28.454536+00:00 · methodology

0 comments
read the original abstract

Rail track extraction is essential for effective railway asset management and maintenance, especially in automated inspection and mapping workflows. This paper introduces a novel method for extracting rail tracks from classified 3D point clouds using a fully convolutional recurrent neural network that preserves full spatial resolution and is trained exclusively on synthetically generated data. This approach enhances per-pixel quality and is particularly suited for rail track extraction. The proposed method begins by rasterizing points corresponding to railroad tracks, then applies the neural network to reduce noise and yield a cleaner track representation suitable for vectorization [1]. Subsequent morphological operations further refine the resultant data, enabling accurate track centerline extraction. Next, the extracted centerlines undergo smoothing to eliminate residual irregularities [2, 3]. Finally, the algorithm transfers 3D information from lidar points onto 2D polylines and applies additional vertical smoothing. A single centerline for both tracks is found using the Dynamic Time Warping (DTW) algorithm [4]. The final outcome consists of rail top centerlines and track centerlines derived for rail pairs, with minimal manual intervention. Experimental validation confirms the effectiveness of this method in yielding high-quality rail track extraction.

Figures

Figures reproduced from arXiv: 2607.06829 by Alexander Gribov, Jie Chang.

Figure 1
Figure 1. Figure 1: Point clouds of a multitrack railway corridor acquired using a train-mounted MLS system. Data courtesy of Esri India. The scene contains typical railway corridor features captured in MLS data, including rails, sleepers, ballast, and ground surfaces, along with vegetation growing between the parallel tracks. Overhead railway electrification infrastructure is also present, consisting of masts and supporting … view at source ↗
Figure 2
Figure 2. Figure 2: Synthetic training examples. The top row shows the original noiseless track rasters; the bottom row shows the corresponding distorted rasters with added noise. Each column pairs the same layout before and after corruption. Validation of neural network performance was conducted primarily using synthetic datasets, as substan￾tial volumes of synthetic data allowed comprehensive training and comparative analys… view at source ↗
Figure 3
Figure 3. Figure 3: A custom 2D convolution block. The Conv2d node has Cin input channels and Cout output channels, kernel radius r, padding r · d, dilation d, and no bias. The subsequent BatchNorm2d and ReLU use Cout channels [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: A stacked arrangement of two consecutive CustomConv2d blocks. Here, Cin and Cout denote the number of input and output feature channels, respectively; r is the kernel radius; and d is the dilation factor. The first block uses Cin as input channels and Cout as output channels, while the second block reuses Cout for both its input and output [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: shows a progressively combined block, ProgressiveCombine2d. In this block, an initial Custom￾Conv2d layer with c channels is applied to the input. Its output is concatenated with the original input via a skip connection, forming 2c channels that are then reduced back to c channels by a subsequent CustomConv2d layer. in CustomConv2d in_channels: c out_channels: c kernel_radius: r dilation: d Concat CustomCo… view at source ↗
Figure 6
Figure 6. Figure 6: The network begins with a channel-expansion front end—informally called an “encoder”—that projects the two-channel input into 24 feature maps while preserving full spatial resolution. These 24-channel maps feed into the Full-Resolution Dilated Fusion bridge, which comprises seven dilated-convolution blocks with dilation rates 2 n − 1, n = 1 . . . 7 (i.e. {1, 3, 7, 15, 31, 63, 127}), to sequentially aggrega… view at source ↗
Figure 7
Figure 7. Figure 7: A recurrent architecture with a decision diamond after the FRPDF layer. From the diamond block, the No arrow (exiting left and down) proceeds to the 1 2 threshold and then to the final output, and the Yes arrow (exiting right and up) loops recurrently via the 3 4 threshold, which feeds back to the Concat block [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Example of extracting the railhead top-center polylines (red) for each rail and the resultant track center polyline (blue). Data courtesy of Esri India. 11 Conclusion and Future Work 11.1 Conclusion This paper presents a novel framework for extracting rail tracks and centerlines from classified mobile lidar point clouds using a raster-based deep learning approach. The method combines rasterization, the rec… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

33 extracted references · 33 canonical work pages · 3 internal anchors

  1. [1]

    Vectorization and parity errors,

    A. Gribov and E. Bodansky, “Vectorization and parity errors,” inGraphics Recognition. Ten Years Review and Future Perspectives, W. Liu and J. Lladós, Eds. Springer Berlin Heidelberg, 2006, pp. 1–10. [Online]. Available: https://doi.org/10.1007/11767978_1

    work page doi:10.1007/11767978_1 2006

  2. [2]

    Smoothing and compression of lines obtained by raster-to-vector conversion,

    E. Bodansky, A. Gribov, and M. Pilouk, “Smoothing and compression of lines obtained by raster-to-vector conversion,” inGraphics Recognition Algorithms and Applications, ser. Lecture Notes in Computer Science, D. Blostein and Y.-B. Kwon, Eds., vol. 2390. Springer Berlin Heidelberg, 2002, pp. 256–265. [Online]. Available: https://doi.org/10.1007/3-540-45868...

    work page doi:10.1007/3-540-45868-9_22 2002

  3. [3]

    Smoothing a network of planar polygonal lines obtained with vectorization,

    A. Gribov and E. Bodansky, “Smoothing a network of planar polygonal lines obtained with vectorization,” in Graphics Recognition. Recent Advances and New Opportunities, ser. Lecture Notes in Computer Science, W. Liu, J. Lladós, and J.-M. Ogier, Eds., vol. 5046. Springer Berlin Heidelberg, 2008, pp. 213–234. [Online]. Available: https://doi.org/10.1007/978-...

    work page doi:10.1007/978-3-540-88188-9_21 2008

  4. [4]

    Dynamic time warping algorithm review,

    P. Senin, “Dynamic time warping algorithm review,” 2008. [Online]. Available: https://api.semanticscholar.org/ CorpusID:16629907

    work page 2008

  5. [5]

    Automated extraction of 3-D railway tracks from mobile laser scanning point clouds,

    B. Yang and L. Fang, “Automated extraction of 3-D railway tracks from mobile laser scanning point clouds,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 7, no. 12, pp. 4750–4761, Dec 2014. [Online]. Available: https://doi.org/10.1109/JSTARS.2014.2312378

    work page doi:10.1109/jstars.2014.2312378 2014

  6. [6]

    Positive train control (PTC),

    United States Department of Transportation, “Positive train control (PTC),” retrieved July 11,

    work page

  7. [7]

    Available: https://railroads.dot.gov/research-development/program-areas/train-control/ptc/ positive-train-control-ptc

    [Online]. Available: https://railroads.dot.gov/research-development/program-areas/train-control/ptc/ positive-train-control-ptc

    work page

  8. [8]

    Three emerging LiDAR applications for rail data capture,

    Bartlett & West, “Three emerging LiDAR applications for rail data capture,” retrieved July 11, 2025. [Online]. Available: https://www.bartlettwest.com/insight/three-emerging-lidar-applications-rail-data-capture

    work page 2025

  9. [9]

    A fast algorithm for rail extraction using mobile laser scanning data,

    Y. Lou, T. Zhang, J. Tang, W. Song, Y. Zhang, and L. Chen, “A fast algorithm for rail extraction using mobile laser scanning data,”Remote Sensing, vol. 10, no. 12, pp. 1–22, 2018, article number: 1998. [Online]. Available: https://doi.org/10.3390/rs10121998

    work page doi:10.3390/rs10121998 2018

  10. [10]

    Semantic segmentation of point clouds with PointNet and KPConv architectures applied to railway tunnels,

    M. Soilán, A. Nóvoa, A. Sánchez-Rodríguez, B. Riveiro, and P. Arias, “Semantic segmentation of point clouds with PointNet and KPConv architectures applied to railway tunnels,”ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, vol. V-2-2020, pp. 281–288, 2020. [Online]. Available: https://doi.org/10.5194/isprs-annals-V-2-2...

    work page doi:10.5194/isprs-annals-v-2-2020-281-2020 2020

  11. [11]

    Point cloud semantic segmentation of complex railway environments using deep learning,

    J. Grandio, B. Riveiro, M. Soilán, and P. Arias, “Point cloud semantic segmentation of complex railway environments using deep learning,”Automation in Construction, vol. 141, pp. 1–18, 2022, article number: 104425. [Online]. Available: https://doi.org/10.1016/j.autcon.2022.104425

    work page doi:10.1016/j.autcon.2022.104425 2022

  12. [12]

    Multimodal deep learning for point cloud panoptic segmentation of railway environments,

    J. Grandio, B. Riveiro, D. Lamas, and P. Arias, “Multimodal deep learning for point cloud panoptic segmentation of railway environments,”Automation in Construction, vol. 150, pp. 1–12, 2023, article number: 104854. [Online]. Available: https://doi.org/10.1016/j.autcon.2023.104854

    work page doi:10.1016/j.autcon.2023.104854 2023

  13. [13]

    Automatic extraction of railroad centerlines from mobile laser scanning data,

    S. O. Elberink and K. Khoshelham, “Automatic extraction of railroad centerlines from mobile laser scanning data,”Remote Sensing, vol. 7, no. 5, pp. 5565–5583, 2015. [Online]. Available: https://www.mdpi.com/2072-4292/7/5/5565

    work page 2015

  14. [14]

    Rail track detection and projection-based 3D modeling from UAV point cloud,

    S. Sahebdivani, H. Arefi, and M. Maboudi, “Rail track detection and projection-based 3D modeling from UAV point cloud,”Sensors, vol. 20, no. 18, pp. 1–15, 2020, article number: 5220. [Online]. Available: https://doi.org/10.3390/s20185220

    work page doi:10.3390/s20185220 2020

  15. [15]

    Fully automated methodology for the delineation of railway lanes and the generation of IFC alignment models using 3D point cloud data,

    M. Soilán, A. Nóvoa, A. Sánchez-Rodríguez, A. Justo, and B. Riveiro, “Fully automated methodology for the delineation of railway lanes and the generation of IFC alignment models using 3D point cloud data,”Automation in Construction, vol. 126, pp. 1–21, 2021, article number: 103684. [Online]. Available: https://doi.org/10.1016/j.autcon.2021.103684

    work page doi:10.1016/j.autcon.2021.103684 2021

  16. [16]

    Fully automated extraction of railtop centerline from mobile laser scanning data,

    A. Kononen, H. Kaartinen, A. Kukko, M. Lehtomäki, J. Taher, and J. Hyyppä, “Fully automated extraction of railtop centerline from mobile laser scanning data,”Automation in Construction, vol. 168, pp. 1–10, 2024, article number: 105812. [Online]. Available: https://doi.org/10.1016/j.autcon.2024.105812

    work page doi:10.1016/j.autcon.2024.105812 2024

  17. [17]

    ACNN: a Full Resolution DCNN for Medical Image Segmentation

    X.-Y. Zhou, J.-Q. Zheng, P. Li, and G.-Z. Yang, “ACNN: a full resolution DCNN for medical image segmentation,” 2020. [Online]. Available: https://arxiv.org/abs/1901.09203

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  18. [18]

    Classify rail points in mobile lidar point clouds,

    J. Chang and A. Gribov, “Classify rail points in mobile lidar point clouds,” retrieved May 14, 2025. [Online]. Avail- able: https://www.esri.com/arcgis-blog/products/3d-gis/3d-gis/classify-rail-points-in-mobile-lidar-point-clouds

    work page 2025

  19. [19]

    Prepare point cloud training data (3D Analyst),

    “Prepare point cloud training data (3D Analyst),” retrieved February 4, 2026. [Online]. Available: pro.arcgis.com/en/pro-app/3.4/tool-reference/3d-analyst/prepare-point-cloud-training-data.htm

    work page 2026

  20. [20]

    Train point cloud classification model (3D Analyst),

    “Train point cloud classification model (3D Analyst),” retrieved February 4, 2026. [Online]. Available: pro.arcgis.com/en/pro-app/3.4/tool-reference/3d-analyst/train-point-cloud-classification-model.htm

    work page 2026

  21. [21]

    Classify point cloud using trained model (3D Analyst),

    “Classify point cloud using trained model (3D Analyst),” retrieved February 4, 2026. [Online]. Available: pro.arcgis.com/en/pro-app/3.4/tool-reference/3d-analyst/classify-point-cloud-using-trained-model.htm

    work page 2026

  22. [22]

    Association of genomic subtypes of lower-grade gliomas with shape features automatically extracted by a deep learning algorithm,

    M. Buda, A. Saha, and M. A. Mazurowski, “Association of genomic subtypes of lower-grade gliomas with shape features automatically extracted by a deep learning algorithm,”Computers in Biology and Medicine, vol. 109,

    work page

  23. [23]

    Available: https://doi.org/10.1016/j.compbiomed.2019.05.002

    [Online]. Available: https://doi.org/10.1016/j.compbiomed.2019.05.002

    work page doi:10.1016/j.compbiomed.2019.05.002 2019

  24. [24]

    U-Net for brain segmentation,

    mateuszbuda, “U-Net for brain segmentation,” 2019, retrieved January 5, 2024. [Online]. Available: https://github.com/mateuszbuda/brain-segmentation-pytorch

    work page 2019

  25. [25]

    U-Net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional networks for biomedical image segmentation,” inMedical Image Computing and Computer-Assisted Intervention – MICCAI 2015, N. Navab, J. Hornegger, W. M. Wells, and A. F. Frangi, Eds. Cham: Springer International Publishing, 2015, pp. 234–241. [Online]. Available: https://doi.org/10.1007/978-3-31...

    work page doi:10.1007/978-3-319-24574-4_28 2015

  26. [26]

    Multi-Scale Context Aggregation by Dilated Convolutions

    F. Yu and V. Koltun, “Multi-scale context aggregation by dilated convolutions,” 2016. [Online]. Available: https://arxiv.org/abs/1511.07122

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  27. [27]

    ACNN: a full resolution DCNN for medical image segmentation,

    XiaoYunZhou27, “ACNN: a full resolution DCNN for medical image segmentation,” 2020, retrieved May 25,

    work page 2020

  28. [28]

    Available: https://github.com/XiaoYunZhou27/ACNN

    [Online]. Available: https://github.com/XiaoYunZhou27/ACNN

    work page

  29. [29]

    Batch normalization: accelerating deep network training by reducing internal covariate shift,

    S. Ioffe and C. Szegedy, “Batch normalization: accelerating deep network training by reducing internal covariate shift,” inProceedings of the 32nd International Conference on International Conference on Machine Learning - Volume 37, ser. ICML’15. JMLR.org, 2015, pp. 448–456. [Online]. Available: https://static.googleusercontent.com/media/research.google.c...

    work page 2015

  30. [30]

    [Online]

    Environmental Systems Research Institute (ESRI),Using ArcScan for ArcGIS, 2004. [Online]. Available: https://downloads.esri.com/support/documentation/other_/Using_ArcScan_for_ArcGIS.pdf

    work page 2004

  31. [31]

    What is ArcScan?

    Esri, “What is ArcScan?” retrieved March 17, 2025. [Online]. Available: https://desktop.arcgis.com/en/arcmap/ latest/extensions/arcscan/what-is-arcscan-.htm

    work page 2025

  32. [32]

    Efficient Computation of the Directional Extremal Boundary of a Union of Equal-Radius Circles

    A. Gribov, “Efficient computation of the directional extremal boundary of a union of equal-radius circles,” pp. 1–6, March 2025. [Online]. Available: https://arxiv.org/abs/2503.21711

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  33. [33]

    Extract rails from point cloud (3D Analyst),

    “Extract rails from point cloud (3D Analyst),” retrieved February 4, 2026. [Online]. Available: pro.arcgis.com/en/pro-app/latest/tool-reference/3d-analyst/extract-rails-from-point-cloud.htm Appendix: Integral of the Squared Distance from a Line Segment to a Point Let p qbe a line segment of lengthL=∥q−p∥.Given a pointr, we seek the integral of the squared...

    work page 2026