arxiv: 2604.16513 · v1 · submitted 2026-04-15 · 💻 cs.CV · cs.LG

Recognition: unknown

SynthPID: P&ID digitization from Topology-Preserving Synthetic Data

Pinak Mahapatra, Suraj Prasad

Authors on Pith no claims yet

Pith reviewed 2026-05-10 13:00 UTC · model grok-4.3

classification 💻 cs.CV cs.LG

keywords P&ID digitizationsynthetic datatopology preservationprocess graphsedge detectionRelationformercomputer visiondiagram analysis

0 comments

The pith

Synthetic P&IDs seeded with real pipe topologies train a model to 63.8% edge mAP on real benchmarks without using any real training images.

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

The paper tackles the scarcity of public P&ID data for training digitization models by generating synthetic diagrams that copy the actual pipe connection structures found in real plant drawings. Earlier synthetic methods placed symbols at random and reached only around 33% edge accuracy, but the new corpus of 665 topology-seeded images allows a patch-based Relationformer trained exclusively on them to hit 63.8 +/- 3.1% edge mAP on the PID2Graph OPEN100 test set. This result sits just 8 points below a model trained on real data and holds in direct comparisons that isolate generation quality from model choice. The work shows that matching real topologies, rather than visual style alone, is what closes most of the domain gap for this task.

Core claim

SynthPID is a corpus of 665 synthetic P&IDs created by seeding pipe topologies directly from real drawings. A patch-based Relationformer trained solely on this corpus reaches 63.8 +/- 3.1% edge mAP on the OPEN100 benchmark without seeing any real P&IDs, closing within 8 points of the real-data oracle. Gains saturate beyond roughly 400 synthetic images, indicating that seed diversity rather than volume is the binding constraint.

What carries the argument

SynthPID, the corpus of 665 synthetic P&IDs whose pipe topologies are seeded directly from real drawings, which is paired with a patch-based Relationformer adapted for high-resolution diagrams.

If this is right

A model trained only on synthetic data can reach performance close to real-data training for extracting process graphs from P&IDs.
Preserving real pipe topologies during synthesis is essential, as random symbol placement yields far lower accuracy.
Performance improves with added synthetic images but levels off around 400, limited by the variety of available real seed topologies.
Generation quality, not network architecture, accounts for the observed gains in this controlled comparison.

Where Pith is reading between the lines

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

Collecting more distinct real pipe topologies as seeds could further narrow the remaining 8-point gap to real-data performance.
The same seeding approach might extend to digitizing other proprietary engineering documents that share topological structure.
If the 100-image benchmark underrepresents industrial variability, additional real test sets would be needed to confirm broad applicability.

Load-bearing premise

Synthetic P&IDs seeded with real pipe topologies are distributed closely enough to real P&IDs for models to generalize without large domain shift, and the single public benchmark of 100 images is representative of typical plant drawings.

What would settle it

Evaluating the SynthPID-trained model on a larger or topologically distinct collection of real P&IDs and measuring a drop below 50% edge mAP would indicate that the synthetic distribution fails to capture real variability.

Figures

Figures reproduced from arXiv: 2604.16513 by Pinak Mahapatra, Suraj Prasad.

**Figure 1.** Figure 1: Method overview. Twelve real P&IDs serve as structural seeds. Our generator perturbs each seed’s topology and re-renders it, producing SynthPID: 665 diverse synthetic P&IDs with full graph annotation. After connector collapsing and patch extraction, a Relationformer is trained on these patches and evaluated on the original real images via patch merging. perturbing it, rather than building from an empty can… view at source ↗

**Figure 2.** Figure 2: The visual domain gap. Real OPEN100 P&IDs (top) have white backgrounds and professional drafting conventions; SynthPID images (bottom) have a grey background and a different rendering style. Despite this, both share the same symbol vocabulary and, critically, the same pipe topology statistics. 1500 × 1500 px patches at a stride of 750 px, which brings individual symbols to 50-200 px, a comfortable range f… view at source ↗

**Figure 3.** Figure 3: Degree distribution (a) and edge density (b) for three corpora. SynthPID tracks the real OPEN100 distribution closely; template-based synthetic data does not. patch duplicate suppression via NMS followed by Weighted Box Fusion (WBF) [17], border-node matching to reconnect cross-boundary edges, and removal of self-loops and isolated nodes. Both SynthPID and OPEN100 pass through this identical pipeline, so … view at source ↗

**Figure 4.** Figure 4: Edge mAP versus number of synthetic training images. Performance rises sharply up to around 400 images (+12.5 pp from 100) and then begins to level off (+5.1 pp from 400 to 665), suggesting that seed diversity rather than image volume is the limiting factor. 4.4. Per-Class Analysis [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗

**Figure 5.** Figure 5: Ground truth (left) versus SynthPID predictions (right) on a real OPEN100 P&ID (top row) and a synthetic image (bottom row), both under E2. Solid boxes and solid edge lines are true positives; dashed elements are false positives. The model reconstructs the main process topology without any real training data, with errors concentrated in the general class and densely packed regions [PITH_FULL_IMAGE:figures… view at source ↗

read the original abstract

Automating the digitization of Piping and Instrumentation Diagrams (P&IDs) into structured process graphs would unlock significant value in plant operations, yet progress is bottlenecked by a fundamental data problem: engineering drawings are proprietary, and the entire community shares a single public benchmark of just 12 annotated images. Prior attempts at synthetic augmentation have fallen short because template-based generators scatter symbols at random, producing graphs that bear little resemblance to real process plants and, accordingly, yield only approximately 33% edge detection accuracy under synth-only training. We argue the failure is structural rather than visual and address it by introducing SynthPID, a corpus of 665 synthetic P&IDs whose pipe topology is seeded directly from real drawings. Paired with a patch-based Relationformer adapted for high-resolution diagrams, a model trained on SynthPID alone achieves 63.8 +/- 3.1% edge mAP on PID2Graph OPEN100 without seeing a single real P&ID during training, closing within 8 pp of the real-data oracle. These gains hold up under a controlled comparison against the template-based regime, confirming that generation quality drives performance rather than model choice. A scaling study reveals that gains flatten beyond roughly 400 synthetic images, pointing to seed diversity as the binding constraint.

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.

Desk Editor's Note private letter to a colleague

Topology seeding from real drawings lifts synthetic-only training to 63.8% edge mAP on the real PID2Graph benchmark, but the paper supplies no direct checks that the seeded graphs match real distributions.

read the letter

The main takeaway is that seeding synthetic P&IDs with actual pipe topologies extracted from real drawings produces a training set that lets a model reach 63.8 +/- 3.1% edge mAP on the 100-image real test benchmark without ever seeing real training data. This is a clear step up from the 33% reported for template-based random placement, and the gap to the real-data oracle shrinks to 8 points. The controlled comparison and the scaling curve that flattens after roughly 400 images are the parts that feel solid; they tie the gain to the topology choice rather than model tweaks or sheer volume.

Referee Report

2 major / 2 minor

Summary. The paper introduces SynthPID, a dataset of 665 synthetic P&IDs whose pipe topologies are seeded from real drawings, to overcome the scarcity of public annotated P&ID data (only 12 images in prior benchmarks). Paired with a patch-based Relationformer adapted for high-resolution inputs, a model trained exclusively on SynthPID achieves 63.8 ± 3.1% edge mAP on the external PID2Graph OPEN100 benchmark—within 8 pp of a real-data oracle—while template-based synthetics reach only ~33%. A scaling study shows gains plateau after ~400 images, attributing the limit to seed diversity.

Significance. If the central result holds, the work meaningfully advances automated P&ID digitization by showing that topology-preserving synthetic data can substantially close the domain gap without any real training images. The controlled comparison against template-based generation and the scaling curve with error bars provide concrete evidence that data quality, rather than model architecture alone, drives the improvement. These elements, together with evaluation on an external public benchmark, strengthen the contribution relative to prior synthetic-augmentation attempts.

major comments (2)

[§3 (Data Generation)] §3 (Data Generation): The headline claim that topology-seeded synthetics enable 63.8% edge mAP without domain shift rests on the unverified assumption that the 665 generated graphs match the distribution of real P&IDs in OPEN100. No graph-level statistics (e.g., degree distributions, symbol co-occurrence matrices, or topological invariants such as cycle counts) are reported to quantify this match, leaving open the possibility that performance reflects memorization of the (unspecified number of) seed topologies rather than generalization.
[§4 (Experiments)] §4 (Experiments) and Abstract: The reported 8 pp gap to the real-data oracle and the 63.8 ± 3.1% figure are load-bearing for the central contribution, yet the manuscript supplies insufficient detail on the seeding procedure (exact number of distinct real topologies, selection criteria, and explicit confirmation that no test-set topologies appear in the seeds) and on the patch-based Relationformer adaptation (patch size, overlap strategy, and score aggregation). These omissions directly affect reproducibility and interpretation of the controlled comparison.

minor comments (2)

[Abstract] Abstract: The template-based baseline is described only as “approximately 33%”; reporting the exact value with standard deviation (consistent with the SynthPID result) would allow direct quantitative comparison.
[Scaling study] Scaling study: The statement that gains “flatten beyond roughly 400” would be clearer if accompanied by a table or figure showing mAP versus number of synthetic images with error bars for all points.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback and for recognizing the significance of the topology-preserving synthetic data approach. We address each major comment point by point below, agreeing where revisions are warranted to improve the manuscript.

read point-by-point responses

Referee: [§3 (Data Generation)] The headline claim that topology-seeded synthetics enable 63.8% edge mAP without domain shift rests on the unverified assumption that the 665 generated graphs match the distribution of real P&IDs in OPEN100. No graph-level statistics (e.g., degree distributions, symbol co-occurrence matrices, or topological invariants such as cycle counts) are reported to quantify this match, leaving open the possibility that performance reflects memorization of the (unspecified number of) seed topologies rather than generalization.

Authors: We agree that the manuscript would be strengthened by explicit quantitative verification of distributional match. The seeding procedure ensures topological fidelity by construction, as each synthetic P&ID is generated from a real seed graph rather than random templates; this is the core distinction from prior work that achieved only ~33% edge mAP. Nevertheless, we did not report comparative graph statistics in the original submission. In the revision we will add these to §3, including degree distributions, symbol co-occurrence matrices, and cycle counts for SynthPID versus OPEN100, along with the number of distinct seed topologies and evidence that performance scales with seed diversity rather than memorization of a small set. revision: yes
Referee: [§4 (Experiments)] and Abstract: The reported 8 pp gap to the real-data oracle and the 63.8 ± 3.1% figure are load-bearing for the central contribution, yet the manuscript supplies insufficient detail on the seeding procedure (exact number of distinct real topologies, selection criteria, and explicit confirmation that no test-set topologies appear in the seeds) and on the patch-based Relationformer adaptation (patch size, overlap strategy, and score aggregation). These omissions directly affect reproducibility and interpretation of the controlled comparison.

Authors: We acknowledge that the original manuscript omitted sufficient implementation details for full reproducibility. The seeding and model adaptation descriptions were kept concise due to length constraints. We will revise §3 and §4 (and add an appendix if needed) to specify the seeding procedure, including the exact number of distinct real topologies, selection criteria, and explicit confirmation that no test-set topologies from OPEN100 appear among the seeds. We will likewise detail the patch-based Relationformer adaptation, including patch size, overlap strategy, and score aggregation method. These additions will not change the reported results but will allow readers to replicate the controlled comparison. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical result on external real benchmark

full rationale

The paper reports an empirical mAP of 63.8% on the external PID2Graph OPEN100 benchmark after training solely on the 665 SynthPID images. No equations, fitted parameters, or derivations are presented that reduce this metric to a quantity defined by the method itself. The synthetic data seeds topologies from real drawings but the test set consists of held-out real images, so the result is independently falsifiable. No self-citations are invoked as load-bearing uniqueness theorems or ansatzes for the core claim. The scaling study and controlled comparison to template-based synthetics are also external measurements rather than self-referential.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that topology seeding produces synthetic data whose distribution is close enough to real P&IDs to enable generalization; no free parameters are fitted to the target metric and no new entities are postulated.

axioms (1)

domain assumption Synthetic P&IDs generated by seeding pipe topology from real drawings have a distribution sufficiently close to real P&IDs for the model to generalize.
This assumption is required for the synth-to-real transfer result to hold and is not independently verified in the abstract.

pith-pipeline@v0.9.0 · 5526 in / 1399 out tokens · 30859 ms · 2026-05-10T13:00:51.857153+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

22 extracted references · 6 canonical work pages · 1 internal anchor

[1]

Solving rubik’s cube with a robot hand,

Ilge Akkaya et al. Solving Rubik’s Cube with a robot hand. arXiv:1910.07113, 2019. 2

work page arXiv 1910
[2]

Character region awareness for text detection

Youngmin Baek et al. Character region awareness for text detection. InProc. IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), 2019. 2

2019
[3]

Automated symbol detection in p&IDs

Jaehyun Cha et al. Automated symbol detection in p&IDs. In Proc. CVPR Workshops, 2019. 2

2019
[4]

Symbols in engineering drawings: GAN-based data augmentation

Eyad Elyan, Carlos Francisco Moreno-Garcia, and Paul John- ston. Symbols in engineering drawings: GAN-based data augmentation. InProc. Int. Joint Conf. Neural Networks (IJCNN), 2020. 2

2020
[5]

OPEN100: Open-source nuclear reac- tor design.https://www.open-100.com, 2020

Energy Impact Center. OPEN100: Open-source nuclear reac- tor design.https://www.open-100.com, 2020. 4

2020
[6]

Domain-adversarial training of neural networks.Journal of Machine Learning Research, 17(59): 1–35, 2016

Yaroslav Ganin et al. Domain-adversarial training of neural networks.Journal of Machine Learning Research, 17(59): 1–35, 2016. 2

2016
[7]

EGTR: Extracting graph from trans- former for scene graph generation

Donghwa Im et al. EGTR: Extracting graph from trans- former for scene graph generation. InProc. IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), 2024. arXiv:2404.02072. 2

work page arXiv 2024
[8]

Deep learning for engineering diagram digitization: A survey

Peter Jamieson et al. Deep learning for engineering diagram digitization: A survey. InProc. CVPR Workshops, 2024. 2

2024
[9]

phash: The open source perceptual hash library

JPhash. phash: The open source perceptual hash library. https://www.phash.org, 2010. 3

2010
[10]

Harold W. Kuhn. The Hungarian method for the assignment problem.Naval Research Logistics Quarterly, 2:83–97, 1955. 4

1955
[11]

Automatic recognition of P&ID sym- bols.Engineering Applications of Artificial Intelligence, 95: 103916, 2020

Ankita Mani et al. Automatic recognition of P&ID sym- bols.Engineering Applications of Artificial Intelligence, 95: 103916, 2020. 2

2020
[12]

Symbol detection in p&ID diagrams

Jouni Nurminen et al. Symbol detection in p&ID diagrams. In Proc. Int. Conf. Document Analysis and Recognition (ICDAR),
[13]

DigitizePID: Automatic digitization of piping and instrumentation diagrams

Shruti Paliwal et al. DigitizePID: Automatic digitization of piping and instrumentation diagrams. InProc. AAAI Work- shop on Graphs and More Complex Structures for Learning and Reasoning, 2021. 1, 2

2021
[14]

Automatic digitization of engineering diagrams

Anand Rahul et al. Automatic digitization of engineering diagrams. InProc. CVPR Workshops, 2019. 2

2019
[15]

Weisfeiler-lehman graph kernels

Nino Shervashidze et al. Weisfeiler-lehman graph kernels. Journal of Machine Learning Research, 12:2539–2561, 2011. 3

2011
[16]

Relationformer: A unified framework for image-to-graph generation

Suprosanna Shit et al. Relationformer: A unified framework for image-to-graph generation. InProc. European Conf. Com- puter Vision (ECCV), 2022. arXiv:2203.10202. 1, 2, 4

work page arXiv 2022
[17]

Weighted boxes fusion: Ensembling boxes from different object detection models

Roman Solovyev, Weimin Wang, and Tatiana Gabruseva. Weighted boxes fusion: Ensembling boxes from different object detection models. InImage and Vision Computing, page 104117, 2021. 4

2021
[18]

From engineering diagrams to graphs: Digitizing P&IDs with transformers

Jan Marius St ¨urmer, Marius Graumann, and Tobias Koch. From engineering diagrams to graphs: Digitizing P&IDs with transformers. InProc. IEEE Int. Conf. Data Science and Advanced Analytics (DSAA), 2025. arXiv:2411.13929. 1, 2, 3, 4, 5

work page arXiv 2025
[19]

Automatic generation of simulation models from digitized engineering diagrams

Jan Marius St ¨urmer et al. Automatic generation of simulation models from digitized engineering diagrams. arXiv:2311.12670, 2023. 2

work page arXiv 2023
[20]

Domain randomization for transferring deep neural networks from simulation to the real world

Josh Tobin et al. Domain randomization for transferring deep neural networks from simulation to the real world. InProc. IEEE/RSJ Int. Conf. Intelligent Robots and Systems (IROS),
[21]

Piping and instrumentation diagram (P&ID) development, 2019

Alireza Toghraei. Piping and instrumentation diagram (P&ID) development, 2019. 1

2019
[22]

Deformable DETR: Deformable Transformers for End-to-End Object Detection

Xizhou Zhu et al. Deformable DETR: Deformable trans- formers for end-to-end object detection. InProc. Int. Conf. Learning Representations (ICLR), 2021. arXiv:2010.04159. 2, 4

work page internal anchor Pith review arXiv 2021