pith. sign in

arxiv: 2512.07269 · v2 · submitted 2025-12-08 · 💻 cs.CV · cs.LG

A graph generation pipeline for critical infrastructures based on heuristics, images and depth data

Mike Diessner , Yannick E. Tarant This is my paper

Pith reviewed 2026-05-17 00:23 UTC · model grok-4.3

classification 💻 cs.CV cs.LG
keywords graph generationcritical infrastructuresphotogrammetryobject detectionheuristicsdigital twinshydraulic systemsstereo camera
0
0 comments X

The pith

A pipeline using stereo camera images, depth data, and user-defined heuristics generates graphs close to ground truth for critical infrastructures.

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

The paper introduces a graph generation pipeline for modeling critical infrastructures like hydraulic systems. Instead of relying on expensive laser scanners for 3D point clouds, it uses affordable stereo cameras to capture RGB images and depth data. Deep learning models detect and segment the objects in these images. User-defined heuristics then determine the relations between these objects to build the graph. Tests on two hydraulic systems demonstrate that the resulting graphs closely match the actual ground truth configurations, offering a transparent method suitable for high-stakes applications.

Core claim

The pipeline detects relevant objects using deep learning on RGB and depth data from a stereo camera and infers their relations via user-defined heuristics, resulting in graphs that closely approximate the ground truth for two tested hydraulic systems.

What carries the argument

The combination of deep learning-based object detection and instance segmentation with user-defined heuristics for inferring relations between objects from image and depth data.

If this is right

  • Virtual representations of infrastructures can be created more affordably for simulations and digital twins.
  • The explicit rules provide transparency that qualifies the graphs for use in high-stakes critical infrastructure decisions.
  • The general process can be tailored to other types of infrastructures beyond hydraulic systems.
  • Specialist knowledge and equipment for laser scanning are not required.

Where Pith is reading between the lines

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

  • Smaller facilities without access to laser scanners could more readily build simulation models for resilience planning.
  • The method might support extension to dynamic updates if paired with continuous camera feeds.
  • Similar heuristic-driven pipelines could apply to graph construction in adjacent domains like facility management or network topology.

Load-bearing premise

User-defined heuristics can accurately and reliably infer relations between detected objects across varied real-world conditions without extensive additional validation.

What would settle it

Applying the pipeline to images from a third hydraulic system or different infrastructure type and obtaining graphs with low similarity to independently verified ground truth structures.

Figures

Figures reproduced from arXiv: 2512.07269 by Mike Diessner, Yannick E. Tarant.

Figure 1
Figure 1. Figure 1: Synthetic hydraulic systems. (A) shows system 1 consisting of pipes, one pump, one tank, one valve and one sprinkler. (B), (C) and (D) show system 2 consisting of pipes, two pumps, two tanks, three valves and four sprinklers. security and well-being [Lampropoulos et al., 2024, Sousa et al., 2021]. In many cases, a digital twin of a physical asset, such as a nuclear facility or a water treatment plant, requ… view at source ↗
Figure 2
Figure 2. Figure 2: Flowchart of the graph generation pipeline. Pipe and non-pipe objects are treated differently and are combined [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Object detection of pumps, tanks and valves including one keypoint for tanks and two keypoints for pumps and [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Object matching across individual images. ( [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Instance segmentation of pipe objects. Bounding boxes indicate detected objects with class label and class [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Cleanup of pipe segmentations projected to world view. ( [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Projection of segmented pipes. (A) is the target image on which segmented pipes of the source image (B) are projected. (C) shows the projected segmentations from (B) in color on the segmentations of (A) in gray. 7 [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Endpoint approximation of pipe elements. Blue lines are a wireframe representation of the pipe elements [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Graph generation. (A) shows the initial graph based on distance between endpoints, and (B) through (H) show graphs enforced by the individual rules discussed in Section 2.2.3. 9 [PITH_FULL_IMAGE:figures/full_fig_p009_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Comparison of ground truth (A) to predicted graph (B) for hydraulic system 1. 3.1 Hydraulic system 1 Hydraulic system 1 consists of one tank, one pump, one valve and one sprinkler as well as three T-fittings as displayed in the ground truth graph in [PITH_FULL_IMAGE:figures/full_fig_p011_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Comparison of ground truth (A) to predicted graph (B) for hydraulic system 2. tanks is near perfect with only a minor difference in the number of predicted pipe elements between the valves and the pumps and between the pump and the T-fitting of the longer arm. This difference is due to the detection of extra pipe elements by YOLOv8 [Redmon et al., 2016], as discussed in Section 2.2.2, and the consequent i… view at source ↗
Figure 12
Figure 12. Figure 12: Prediction of individual objects for hydraulic system 1 ( [PITH_FULL_IMAGE:figures/full_fig_p013_12.png] view at source ↗
read the original abstract

Virtual representations of physical critical infrastructures, such as water or energy plants, are used for simulations and digital twins to ensure resilience and continuity of their services. These models usually require 3D point clouds from laser scanners that are expensive to acquire and require specialist knowledge to use. In this article, we present a prototypical graph generation pipeline based on photogrammetry. The pipeline detects relevant objects and predicts their relation using RGB images and depth data generated by a stereo camera. This more cost-effective approach uses deep learning for object detection and instance segmentation of the objects, and employs user-defined heuristics or rules to infer their relations. Results of two hydraulic systems show that this strategy can produce graphs close to the ground truth. While this study focuses on hydraulic systems, the general process can be used to tailor the method to other types of infrastructures and applications. The user-defined rules create transparency qualifying the pipeline to be used in the high stakes decision-making that is required for critical infrastructures.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript presents a prototypical pipeline for generating graph-based virtual representations of critical infrastructures (e.g., hydraulic systems) from RGB images and depth data captured by a stereo camera. Deep learning performs object detection and instance segmentation, while user-defined heuristics infer object relations to construct the final graph. The authors report that the approach yields graphs close to ground truth on two hydraulic systems and argue that the transparent, rule-based design makes it suitable for high-stakes critical-infrastructure applications as a lower-cost alternative to laser scanning.

Significance. If the central claim were supported by quantitative evidence, the work would offer a practical, interpretable method for creating digital twins of critical systems without specialized 3D scanning equipment. The explicit use of user-defined rules is a positive feature for auditability in regulated domains. However, the absence of any numerical evaluation metrics or validation procedures substantially reduces the current significance and prevents assessment of whether the pipeline generalizes beyond the two reported examples.

major comments (3)
  1. [Abstract] Abstract: The assertion that 'Results of two hydraulic systems show that this strategy can produce graphs close to the ground truth' is unsupported by any quantitative graph-level metrics (e.g., edge precision/recall, graph-edit distance, or structural similarity scores), error analysis, or dataset statistics, rendering the central empirical claim impossible to evaluate.
  2. [Results] Results / Evaluation section: With success reported on exactly two systems and no ablation studies, cross-condition tests, or comparison against alternative relation-inference methods, it is unclear whether the observed closeness to ground truth stems from the pipeline itself or from manual tuning of the system-specific heuristics.
  3. [Methods] Methods: The user-defined heuristics for relation inference are presented as a core component, yet no formal description, pseudocode, or sensitivity analysis is supplied; this makes reproducibility and extension to other infrastructure types difficult to assess.
minor comments (2)
  1. Clarify how depth data from the stereo camera is integrated with the photogrammetry pipeline and whether any 3D reconstruction step occurs before graph construction.
  2. [Abstract] The abstract states the method 'can be used to tailor the method to other types of infrastructures'; provide at least one concrete example or discussion of the required changes to the heuristic set.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their detailed review and constructive comments. We address each major comment below and indicate the revisions we will make to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The assertion that 'Results of two hydraulic systems show that this strategy can produce graphs close to the ground truth' is unsupported by any quantitative graph-level metrics (e.g., edge precision/recall, graph-edit distance, or structural similarity scores), error analysis, or dataset statistics, rendering the central empirical claim impossible to evaluate.

    Authors: We agree that the claim in the abstract would benefit from quantitative support. In the revised manuscript, we will include specific graph-level metrics such as edge precision, recall, and graph-edit distance for the two hydraulic systems, as well as dataset statistics and a brief error analysis. revision: yes

  2. Referee: [Results] Results / Evaluation section: With success reported on exactly two systems and no ablation studies, cross-condition tests, or comparison against alternative relation-inference methods, it is unclear whether the observed closeness to ground truth stems from the pipeline itself or from manual tuning of the system-specific heuristics.

    Authors: This work is a prototypical demonstration on two systems. We acknowledge the limited evaluation scope. In the revision, we will add a discussion clarifying the heuristic development process based on domain expertise and include a sensitivity analysis to address concerns about manual tuning. revision: partial

  3. Referee: [Methods] Methods: The user-defined heuristics for relation inference are presented as a core component, yet no formal description, pseudocode, or sensitivity analysis is supplied; this makes reproducibility and extension to other infrastructure types difficult to assess.

    Authors: We agree on the need for a formal description to aid reproducibility. The revised manuscript will include pseudocode for the relation inference heuristics and a sensitivity analysis on the key parameters. revision: yes

Circularity Check

0 steps flagged

No circularity: pipeline relies on external DL models and explicit user heuristics with no derivations or self-referential predictions

full rationale

The paper describes an applied pipeline for graph generation from RGB/depth images: deep learning handles object detection and instance segmentation, while user-defined heuristics infer relations between objects. No equations, fitted parameters, predictions, or derivation steps are presented that could reduce to inputs by construction. Results on two hydraulic systems are empirical demonstrations of the pipeline output, not derived claims that loop back to the same data or rules. No self-citations are invoked as load-bearing uniqueness theorems or ansatzes. This matches the default expectation of no circularity for a methods paper without mathematical modeling.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The approach depends on pre-existing deep learning models for detection/segmentation and custom user heuristics for relations; no free parameters, axioms, or invented entities are specified in the abstract.

pith-pipeline@v0.9.0 · 5468 in / 1169 out tokens · 70684 ms · 2026-05-17T00:23:05.925445+00:00 · methodology

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

3 extracted references · 3 canonical work pages

  1. [1]

    Qiuchen Lu, Xiang Xie, Ajith Kumar Parlikad, Jennifer Mary Schooling, and Eirini Konstantinou

    doi:10.1007/978-3-319-92862-3_1. Qiuchen Lu, Xiang Xie, Ajith Kumar Parlikad, Jennifer Mary Schooling, and Eirini Konstantinou. Moving from building information models to digital twins for operation and maintenance.Proceedings of the Institution of Civil Engineers-Smart Infrastructure and Construction, 174(2):46–56, 2022. doi:10.1680/jsmic.19.00011. Andrz...

    work page doi:10.1007/978-3-319-92862-3_1 2022

  2. [2]

    Lili Cheng, Zhuo Wei, Mingchao Sun, Shiqing Xin, Andrei Sharf, Yangyan Li, Baoquan Chen, and Changhe Tu

    doi:10.7315/JCDE.2014.002. Lili Cheng, Zhuo Wei, Mingchao Sun, Shiqing Xin, Andrei Sharf, Yangyan Li, Baoquan Chen, and Changhe Tu. DeepPipes: Learning 3D pipelines reconstruction from point clouds.Graphical Models, 111:101079, 2020. doi:10.1016/j.gmod.2020.101079. Antje Alex and Jannis Stoppe. Pipe reconstruction from point cloud data. InProceedings of t...

    work page doi:10.7315/jcde.2014.002 2014

  3. [3]

    Feiyu Xu, Hans Uszkoreit, Yangzhou Du, Wei Fan, Dongyan Zhao, and Jun Zhu

    doi:10.1007/s00521-024-10437-2. Feiyu Xu, Hans Uszkoreit, Yangzhou Du, Wei Fan, Dongyan Zhao, and Jun Zhu. Explainable AI: A brief survey on history, research areas, approaches and challenges. InProceedings of the CFF International Conference on Natural 17 Graph generation of critical infrastructureA PREPRINT Language Processing and Chinese Computing (NLP...

    work page doi:10.1007/s00521-024-10437-2 2019