acceptodds
Under review as a conference paper at ICLR 2027

TopoPID: Residual-Guided Equipment Recognition and Graph Completion in P&IDs

Abstract

Digitizing a piping and instrumentation diagram (P&ID) requires an auditable process graph, not just detected symbols. Project-specific equipment often lacks visual exemplars: closed-vocabulary pipelines preserve structure but miss such devices, whereas whole-page vision–language models (VLMs) face dense visual search and cannot directly authorize graph edits. We formulate recovery as residual-guided constrained semantic completion of a partially observed graph. In TopoPID, unresolved topology localizes equipment-scale candidates, a project roster constrains local VLM verification, and deterministic rules govern evidence-linked graph updates. On 45 held-out industrial P&IDs, with target-equipment images excluded from detector development, prompt demonstrations, and candidate tuning, identity-aware recognition F1 rises from 19.44% for whole-page inference to 93.77%; 489/524 devices are correctly recovered. Without adaptation, the same configuration recovers 54/62 task-compatible devices (86.40% F1) on 12 OPEN100 drawings. These results establish device recovery and useful graph association retrieval while retaining an auditable write-back path.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
piping and instrumentation diagrams, graph completion, equipment recognition, vision-language models, engineering drawings

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_M1GCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents TopoPID, a framework for recognizing previously unseen equipment in P&IDs and completing partial process graphs. It combines topology-residual localization, roster-grounded VLM verification, and deterministic graph updates with provenance tracking. Experiments on 45 industrial drawings achieve 93.77% identity-aware F1, outperforming whole-page inference (19.44%) and endpoint crops (80.38%). The method further achieves 86.40% F1 on OPEN100 and 95.58% instrument–equipment association retrieval accuracy.

Critical strengths
  1. The integration of topology-guided localization, roster-constrained semantic verification, and evidence-preserving graph updates provides a practical and interpretable solution to unseen-equipment recognition.

  2. The substantial improvements over fixed-VLM baselines, together with encouraging zero-adaptation performance on OPEN100, demonstrate the effectiveness of structured localization.

Critical weaknesses
  1. The contributions of endpoint-density aggregation, visual residuals, and candidate refinement are not individually evaluated. Comparisons with stronger open-vocabulary detection or alternative proposal strategies are also missing, limiting the assessment of methodological novelty.

  2. Graph completion is evaluated through association retrieval rather than direct edge-level correctness. The absence of a pre-completion baseline and insertion/deletion accuracy makes structural reliability unclear. Unavailable implementation details and the small public evaluation subset further limit reproducibility.

Questions for authors
  1. Can the authors provide component-wise ablations and computational-cost comparisons to isolate the contributions of topology density, visual residuals, and candidate refinement?

  2. How accurately does TopoPID insert and remove graph edges? Could the authors report edge-level precision/recall and association retrieval performance before graph completion?

2 comments

Thank you for your constructive feedback.

1. Structural insight and component contributions.

Our key insight is that an incomplete graph contains evidence about its missing entities: recovered pipes and instruments leave unresolved connectivity that can reveal visually unrecognized equipment. Our ablations examine how this structural evidence improves equipment localization.

Table A. Component comparisons on 524 devices in 45 industrial drawings, using the same semantic verifier.

Proposal methodCandidatesRecognition F1 (%)
Endpoint crops1,86980.38
Density alone1,05175.11
Density + visual residuals (no refinement)1,21684.43
Density + refinement (no visual residuals)1,20982.89
Full TopoPID1,38093.77

Density alone falls below endpoint crops: concentrating endpoint evidence does not necessarily recover a complete device extent. Visual residuals provide complementary spatial support, while refinement addresses the extent and separation of candidate regions. Each addition improves on density alone, and their combination gives the strongest result. Full TopoPID also covers 512/524 devices before verification, versus 421/524 for endpoint crops. These results support broader candidate coverage and more effective localization through complementary structural and visual evidence. Refinement operations, including deduplication, are evaluated jointly.

2. Comparison with appearance-driven localization.

Grounding DINO receives equipment queries drawn from the drawing text. Retaining the highest-scoring box per query gives 12.40% identity-aware F1. Multiple DINO proposals followed by the same Qwen verification settings as TopoPID give 45.02%, compared with 93.77% for TopoPID. Correct recognition requires both IoU > 0.50 and the correct normalized equipment identity.

With a common downstream verifier, this comparison supports the value of using unresolved connectivity to guide localization. The partial graph contributes evidence that text-conditioned visual proposals do not use. The methods retain their respective proposal sets: 977 candidates for DINO+VLM and 1,380 for TopoPID.

3. Direct edge-level accuracy and graph updates.

We agree that association retrieval alone does not establish connection correctness. We therefore evaluate device–endpoint incidences directly on the same 489 correctly recognized equipment regions and 3,443 reference incidences. Correctly matched incidences are true positives, unmatched or duplicate incidences are false positives, and missing reference incidences are false negatives.

Table B. Direct connection evaluation with equipment recognition held fixed.

Connection methodPrecision (%)Recall (%)F1 (%)
Original TopoPID rules66.2086.6475.05
Constraint-guided extension94.5694.9894.77

The extension reduces incorrect incidences from 1,523 to 188 and missing incidences from 460 to 173. Local visual reasoning distinguishes process lines from equipment outlines and ambiguous structures, while graph updates are checked against image-supported candidates and protected relations. The same structural-inference principle thus governs both discovering a missing entity and integrating it into the graph. Section 3.5 describes the extension alongside the original rule-based method.

On the same 1,763 queries, the original rules and extension achieve 95.58% and 95.75% association accuracy, respectively. Despite this small retrieval difference, direct connection quality improves substantially, as different local attachments can lead to the same reachable equipment. Deletion removes only edges exclusive to empty equipment-returning circuits, preserving those shared with valid routes. Table B directly evaluates device–endpoint attachment correctness, while the association results assess retrieval performance on the completed graphs.

4. Public evaluation and implementation details.

We apply the unchanged enterprise configuration to all 12 OPEN100 drawings, including thresholds, candidate budget, and verification prompt. Evaluating all 67 equipment depictions (including five without visible external attachments) yields 59 correct recognitions, 4 false positives, and 8 misses (90.77% F1). Appendix A reconciles this scope with the submitted 62-instance subset (86.40% F1); no new drawings or predictions are introduced.

Sections 3–4 clarify the partial-graph construction, density and residual computation, refinement, verifier inputs, and matching protocol. Candidate and neighborhood images are jointly provided with the drawing-specific roster, and graph updates distinguish observed structure, inferred incidences, and circuit cleanup. The reproducibility statement specifies the contractual restrictions on access to the enterprise data and implementation code.

Following our earlier response, we provide measured computational costs and publicly inspectable OPEN100 results for equipment recognition and graph completion.

1. Computational cost.

The measurements use an Intel Core i7-11700 CPU and an NVIDIA GeForce RTX 3090 GPU (24 GB), with Qwen3.7-Plus accessed via a paid API. Elapsed time includes code execution and API waiting time. Table C reports the average input/output tokens, API fee, and elapsed time per drawing for equipment recognition and the additional constraint-guided connection stage.

Table C. Average computational cost per drawing.

DatasetStageInput tokensOutput tokensAPI cost (CNY)Time (s)
OPEN100Recognition13,5151,6860.032428.9
OPEN100Connection18,1291,8290.040736.2
EnterpriseRecognition38,2584,1210.087647.4
EnterpriseConnection86,3916,7110.181291.2

Summing the two stage averages gives 65.1 s and CNY 0.0731 per OPEN100 drawing, and 138.6 s and CNY 0.2688 per enterprise drawing. These measurements complement the earlier component ablations by quantifying the practical cost of equipment recognition and relational completion. On the enterprise data, the corresponding recognition F1 is 93.77%, and connection F1 is 94.77% on correctly recognized equipment.

2. Public connection evaluation under the unchanged configuration.

The enterprise connection configuration is applied to all 12 OPEN100 drawings without adaptation. Both methods use the same 59 correctly recognized equipment regions, 177 reference incidences, and 315 association queries.

Table D. OPEN100 connections with equipment recognition held fixed.

MethodTP / FP / FNPrecision (%)Recall (%)F1 (%)Query accuracy
Original rules148 / 14 / 2991.3683.6287.32268/315 (85.08%)
Constraint-guided reasoning154 / 8 / 2395.0687.0190.86279/315 (88.57%)

The extension reduces incorrect incidences from 14 to 8 and missing incidences from 29 to 23, while answering 11 additional queries correctly. These results extend the enterprise evidence: interpreting nearby strokes under explicit graph-update constraints improves both local attachments and downstream access to associated equipment. Section 5.4 and Table 5 report the comparison.

3. Publicly available evaluation materials.

The anonymous results repository now provides the complete OPEN100 evaluation over 12 drawings: Grounding DINO Top1, Grounding DINO+VLM, TopoPID recognition, and the two connection methods. The recognition evaluation covers all 67 reference equipment depictions; the connection comparison uses the fixed 59 correctly recognized regions above.

The repository includes reference equipment names, tags and boxes; per-candidate recognition outputs; per-drawing summaries and visualizations; equipment-level connection process views; 376 connection records across both methods; and 630 method–query records containing returned and acceptable equipment. These materials allow the aggregate results to be checked against individual recognition, attachment, and retrieval outcomes. Together with the method and evaluation descriptions in Sections 3–5, they make the evidence for structure-guided recognition and constrained graph completion directly inspectable.

Reviewer_HgCWOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes TopoPID, a framework for recognizing equipment missed by conventional detectors in piping and instrumentation diagrams (P&IDs) and completing the corresponding process graphs. The method uses disconnected line endpoints and visual residuals to generate candidate equipment regions. A pretrained VLM then verifies these candidates against a project-specific equipment roster. Finally, deterministic rules insert accepted equipment and associated connections into the graph.

Critical strengths
  1. Promising experimental results. The method substantially improves recognition performance over whole-page VLM inference and endpoint cropping.

  2. Separating VLM-based recognition from deterministic graph updates is a useful design choice, especially for engineering applications where incorrect connections can have serious consequences.

Critical weaknesses
  1. The motivation is not clear: how many unseen devices will be in one-time recognition? If only a small number, and no guarantee to 100%, which means double-checking manually. What's the purpose of them?

  2. The framework primarily combines existing components, such as YOLOv8. While the resulting system is effective, the paper does not introduce a new learning architecture, training strategy, or clearly generalizable learning formulation. The authors should better explain the broader machine learning insight.

  3. In Table 2, endpoint cropping alone achieves 80.38% F1, suggesting that localizing the VLM input already addresses much of the difficulty. The full framework improves this to 93.77%, but the individual contributions of endpoint-density aggregation, visual residuals, boundary completion, and candidate deduplication are not evaluated separately.

  4. The paper lacks a strong visual-only baseline.

Questions for authors

How to define the unseen device? Is it possible to update the library to recognize all unseen devices?

2 comments

Thank you for asking us to clarify the motivation and methodological insight. We have strengthened these arguments and added component ablations and an open-vocabulary detector comparison.

1. Why recovering omitted equipment matters.

Equipment is an organizing entity in a P&ID: a missed device can leave several otherwise detected pipe and instrument routes without their engineering referent. The task is therefore consequential even when only part of a drawing is unrecognized. Our evaluation contains 524 target devices across 45 industrial drawings, averaging 11.64 per evaluated drawing. Endpoint cropping correctly recognizes 383 devices and misses 141; TopoPID recognizes 489 and misses 35, while reducing false positives from 46 to 30. The purpose is to recover these missing entities and their relationships in a form that can be checked against the source drawing. Engineering review remains part of use; the recognition and connection evaluations measure the quality of the graph supplied for that review.

2. What is the methodological insight?

We agree that TopoPID does not introduce a new visual backbone, detection architecture, or training strategy. However, its contribution is a different perspective on equipment recognition and graph completion. Our central observation is that an incomplete graph contains evidence about its missing entities. When an equipment body is missed, its surrounding pipes and instruments may still be recovered, leaving spatially organized endpoints and unresolved connectivity. These residuals can reveal where an unrecognized entity may exist even when its appearance is unfamiliar. The methodological insight is therefore to use unexplained structural connectivity as evidence for discovering visually unrecognized entities, rather than treating recognition and graph construction as independent sequential tasks. TopoPID uses the partially recovered graph as an active inference resource: unresolved topology guides localization; local visual evidence and a drawing-specific roster support semantic interpretation; evidence-linked rules govern graph updates. The graph is thus not merely the final output of recognition; its incompleteness helps guide subsequent recognition and completion. The connection extension applies the same principle to ambiguous attachments under explicit graph-update constraints.

3. What explains the improvement over endpoint crops?

Table 3 compares the components using the same verifier.

Table A. Component comparisons (F1, %).

Proposal methodF1
Endpoint crops80.38
Density alone75.11
Density + visual residuals (no refinement)84.43
Density + refinement (no visual residuals)82.89
Full TopoPID93.77

Density alone is insufficient; the two additions separately improve the structural proposals and together give the strongest result. Full TopoPID also covers 512/524 devices before verification, versus 421/524 for endpoint crops. These results support both broader candidate coverage and the value of visual support and region refinement. The refinement operations are evaluated jointly, rather than attributing a separate gain to deduplication.

4. Comparison with appearance-driven localization.

Table 2 adds Grounding DINO with equipment queries drawn from the drawing text: Top1 keeps the highest-scoring box per query; DINO+VLM uses multiple proposals and the same Qwen verification settings as TopoPID.

Table B. Appearance-driven comparison (F1, %).

MethodF1
DINO Top112.40
DINO+VLM45.02
TopoPID93.77

Grounding DINO provides text-conditioned visual localization without the partial-graph prior. With a common downstream verifier, the latter comparison supports the value of structure-guided proposals on these drawings. The methods use their respective proposal sets: 977 candidates for DINO+VLM and 1,380 for TopoPID.

5. Definition of target equipment and library expansion.

“Unseen” means the absence of target-equipment visual exemplars from the base detector’s training data, not the absence of the equipment identity from the project roster. The roster supplies candidate identities—names, printed tags, and source-supported descriptions—and can be updated with newly encountered identities. However, adding an identity does not provide the visual examples needed to train or adapt a conventional detector. TopoPID addresses this complementary setting: it uses the updated roster together with local visual evidence and residual process structure to recognize equipment without target-specific detector training. We do not claim to identify arbitrary equipment whose identity is absent from the roster.

As a follow-up to our response on motivation and methodological insight, we provide a public comparison that further examines the value of structure-guided localization.

1. Appearance-driven and structure-guided localization on OPEN100.

We apply the enterprise configurations unchanged to all 12 OPEN100 drawings, using the equipment roster of each corresponding drawing. The evaluation includes all 67 reference equipment depictions. Grounding DINO receives equipment queries from drawing text; its multi-proposal variant and TopoPID use a common downstream semantic verifier.

Table C. Equipment recognition on OPEN100 without dataset-specific adaptation.

MethodCandidatesTP / FP / FNRecognition F1 (%)
Grounding DINO Top1677 / 60 / 6010.45
Grounding DINO+VLM10329 / 17 / 3851.33
TopoPID16959 / 4 / 890.77

With a common verifier, TopoPID correctly recognizes 59 devices compared with 29 for DINO+VLM, while reducing false positives from 17 to 4. The methods retain their respective proposal sets. Together with the enterprise comparison and component ablations in our earlier response, these results support our central observation that an incomplete graph contains evidence about its missing entities. Recovered pipes and instruments provide spatial evidence for equipment that appearance-driven proposals may fail to localize. The contribution is to make this structural evidence an active resource for recognition and completion.

2. Inspectable evidence for the recognition gains.

The anonymous OPEN100 results repository now provides reference equipment names, tags and boxes, per-candidate results, per-drawing summaries, and visualizations for all three methods. The DINO Top1 views pair each queried target with its predicted region; the DINO+VLM and TopoPID views show candidates alongside final outputs. TopoPID's views also display the structural density used for localization.

These materials allow the comparison to be examined at the level of individual equipment regions and their surrounding structure. They complement the aggregate results by showing how the partial graph informs the search for missing equipment and how local visual and roster evidence supports its interpretation.

Reviewer_WT6jOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents TopoPID, a framework for recovering project-specific equipment that conventional detectors miss in P&IDs. It first builds a conservative partial graph. Clusters of unresolved line endpoints, refined by bounded visual cues, are then used to propose equipment-scale regions. Each crop is verified by a VLM against a project equipment roster, and accepted equipment is written into the graph through deterministic rules. On 45 industrial drawings, identity-aware F1 rises from 19.44% for whole-page VLM inference and 80.38% for endpoint crops to 93.77%. On a 62-instance OPEN100 subset, the method reaches 86.40% without adaptation. The completed graphs also answer 95.58% of instrument–equipment association queries correctly.

Critical strengths
  • Using unresolved topology to decide where to look for missing equipment is an interesting idea. It turns the structure that has already been recovered into a spatial search prior, rather than relying only on appearance-based detection or whole-page VLM inference.

  • Leaving an endpoint unresolved rather than adding an unsupported edge, keeping provenance for every decision, and not letting the VLM edit the graph directly are all sensible choices.

Critical weaknesses
  • Graph completion is in the title and is a central part of the method, but it is not directly evaluated. The main metrics measure whether equipment is found and correctly named. The 95.58% association accuracy shows that the completed graph is useful for queries, but it does not tell us whether the inserted incidence edges and removed circuit edges are correct. The same query on the partial graph before completion is also not reported. Section 6.1 acknowledges this limitation, but without such evidence the graph-completion part of the contribution remains largely unverified. PID2Graph OPEN100 provides graph annotations, so after aligning node and edge representations, the authors could report the precision and recall of inserted edges, the correctness of removed edges, and association accuracy before and after completion.

  • It is unclear which proposal components account for the gain over endpoint crops. The pipeline combines endpoint-density aggregation, bounded visual evidence, basin splitting and merging, outline completion and deduplication, and none of these is ablated. Candidate recall is defined in Section 4.2 but reported only for TopoPID. As a result, we cannot tell whether the 13-point F1 gain comes from covering more devices or from giving the verifier better-framed crops. A density-only variant, followed by adding visual evidence and boundary completion step by step, would make the contribution of each mechanism clear.

  • The role of printed identifiers need clarification. The crops in Figure 2 contain readable equipment labels such as “CND-HTR-166”. Reading such labels is legitimate in practice, but it affects what “unseen equipment recognition” means here. The authors should explain how visible text contributes to the identity prediction, ideally with a text-only matching baseline or results on equipment without visible identifiers. The roster size and the similarity among its entries also affect verification difficulty, but neither is reported.

Questions for authors
  • What evidence validates the incidence edges inserted and circuit edges removed by the rules in Sec. 3.5? Can you quantify their correctness, for example using aligned OPEN100 graph annotations, and report association accuracy before and after completion?

  • How much of the F1 gain over endpoint crops in Table 2 comes from higher candidate coverage versus improved crop framing? Can you report candidate recall and final F1 for endpoint crops, density aggregation alone, and variants adding visual evidence and boundary completion?

  • How much does identity prediction rely on readable equipment identifiers, such as those in Figure 2? Can you provide a text-only matching baseline or results on equipment without readable identifiers? Please also report the roster size and similarity among its entries.

2 comments

Thank you for your constructive feedback. We have added direct connection evaluation and component comparisons to clarify the graph-completion results and the source of the recognition gains.

1. Direct connection accuracy and graph updates.

We agree that association retrieval alone does not establish connection correctness. We therefore evaluate device–endpoint incidences directly on the same 489 correctly recognized equipment regions and 3,443 reference incidences from the industrial evaluation. Correctly matched incidences are true positives, unmatched or duplicate incidences are false positives, and missing reference incidences are false negatives.

Table A. Direct connection evaluation with equipment recognition held fixed.

Connection methodPrecision (%)Recall (%)F1 (%)
Original TopoPID rules66.2086.6475.05
Constraint-guided extension94.5694.9894.77

The original rules correctly recover 2,983 reference incidences. The extension increases this to 3,270, while reducing incorrect incidences from 1,523 to 188 and missing incidences from 460 to 173. Local visual reasoning distinguishes process lines from equipment outlines and ambiguous structures, while graph updates are checked against image-supported candidates and protected relations. Semantic reasoning therefore operates within an evidence-preserving graph-update interface. The same structural-inference principle thus governs both discovering a missing entity and integrating it into the graph. Section 3.5 describes the extension alongside the original rule-based method, and Section 5.4 and Table 5 report their direct connection evaluation.

On the same 1,763 instrument–equipment queries, the graphs completed by the original rules and the extension achieve 95.58% and 95.75% association accuracy, respectively. The substantial improvement in direct connection quality is less visible in these retrieval scores because different local attachments can lead to the same reachable equipment. Table A evaluates device–endpoint attachment correctness, while the association results assess retrieval performance on the completed graphs.

2. Component contributions and candidate coverage.

Our key insight is that an incomplete graph contains evidence about its missing entities: recovered pipes and instruments leave unresolved connectivity that can reveal visually unrecognized equipment. The component comparisons examine how this structural evidence becomes an equipment-scale candidate region. All variants below use the same semantic verifier. Correct recognition requires both IoU > 0.50 and the correct normalized equipment identity.

Table B. Component comparisons on 524 devices in 45 industrial drawings. Candidate counts are measured before verification.

Proposal methodCandidatesRecognition F1 (%)
Endpoint crops1,86980.38
Density alone1,05175.11
Density + visual residuals (no refinement)1,21684.43
Density + refinement (no visual residuals)1,20982.89
Full TopoPID1,38093.77

Density alone falls below endpoint crops: concentrating endpoint evidence does not necessarily recover a complete device extent. Visual residuals provide complementary spatial support, while refinement addresses the extent and separation of candidate regions. The two middle variants add visual residuals or refinement separately to density alone. Each addition improves recognition, and their combination gives the strongest result. Refinement operations, including splitting, merging, outline completion, and deduplication, are evaluated jointly.

Candidate recall measures the fraction of reference devices with at least one proposal satisfying IoU > 0.50. Endpoint crops cover 421/524 devices (80.34%), whereas full TopoPID covers 512/524 (97.71%). Full TopoPID also correctly recognizes 489 devices, compared with 383 for endpoint crops. Since endpoint crops cover only 421 devices, at least 68 of TopoPID's correct recoveries lie outside endpoint-crop coverage. This establishes broader candidate coverage as a source of the recognition gain. Together with the component results, it supports the complementary roles of structural localization, visual support, and region refinement. Section 5.2 reports these component comparisons and the candidate-coverage analysis.

Following our earlier response, we provide controlled results on printed characters and the equipment roster, together with direct connection evaluation on OPEN100.

1. Roles of printed characters and the equipment roster.

We keep the same 1,380 candidates and their neighborhoods fixed and report verification outcomes for the same 512 positive enterprise candidates. Each request includes the candidate image and its neighborhood. In the masked conditions, all characters are masked in both images; the roster condition supplies the current drawing's equipment name–tag list.

Table C. Verification outcomes on 512 positive candidates (%).

Image textRosterAcceptedName correctTag correctName + tag correct
RetainedYes98.4496.2995.5195.51
RetainedNo96.096.8473.446.84
MaskedYes95.5184.7779.6979.69
MaskedNo93.556.050.000.00

All columns use the same denominator of 512, including rejected candidates. Identity correctness additionally requires equipment acceptance; the final column is joint name–tag correctness on these candidates.

With image text retained but the roster removed, tag correctness remains 73.44%, while name correctness is 6.84%. With the roster retained and all image characters masked, 408/512 candidates still receive jointly correct names and tags (79.69%), compared with 489/512 (95.51%) when both sources are available. These results distinguish equipment acceptance from engineering identity: printed identifiers support tag recovery, while the drawing-specific roster supplies identity alternatives for interpreting local visual evidence. They support the intended division of inference: unresolved structure guides where to inspect, and visual and roster evidence grounds what the region represents.

Here, “unseen” refers to the absence of target-equipment visual exemplars from the base detector's training data; candidate identities remain available in the drawing-specific roster. Section 5.2 and Appendix B (Table 7) report these controls.

2. Direct connection evaluation on public drawings.

We evaluate the unchanged enterprise connection configuration on all 12 OPEN100 drawings, holding the same 59 correctly recognized equipment regions fixed, with 177 reference incidences and 315 association queries.

Table D. OPEN100 connections with equipment recognition held fixed.

MethodTP / FP / FNPrecision (%)Recall (%)F1 (%)Query accuracy
Original rules148 / 14 / 2991.3683.6287.32268/315 (85.08%)
Constraint-guided reasoning154 / 8 / 2395.0687.0190.86279/315 (88.57%)

Constraint-guided reasoning recovers six additional correct incidences, reduces incorrect incidences from 14 to 8, and answers 11 additional queries correctly. Direct attachment accuracy and retrieval utility improve together under the transferred configuration, supporting the use of local structural interpretation to integrate recognized entities into the graph. Section 5.4 and Table 5 report this comparison.

The anonymous OPEN100 results repository now provides per-drawing recognition and connection results. Its 04_connection folder contains equipment-level process visualizations, comparisons of the two methods, connection records, and returned/reference query answers, allowing individual attachments and their retrieval consequences to be inspected.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response: Core Insight and Summary of Revisions

We thank the reviewers for their constructive feedback. We have submitted a revised manuscript that clarifies the structural-inference contribution and adds evidence addressing the main methodological and empirical questions.

1. The central insight and contribution.

Equipment is an organizing entity in a P&ID: a missed device can leave otherwise recovered pipe and instrument routes without their engineering referent. Our central observation is that an incomplete graph contains evidence about its missing entities. When an equipment body is missed, surrounding pipes and instruments may still be recovered, leaving spatially organized endpoints and unresolved connectivity. TopoPID uses this partial graph as an active inference resource: unresolved topology guides localization; local visual evidence and a drawing-specific roster support semantic interpretation; evidence-linked rules govern graph updates. Structure therefore provides both evidence for discovering omitted entities and constraints on integrating them into the graph. The connection extension applies this principle to ambiguous attachments under explicit graph-update constraints (Section 3.5).

2. What the additional experiments establish.

Table A. Key recognition and connection results (F1, %).

EvaluationComparison methodTopoPID / extension
Enterprise recognitionDINO+VLM: 45.02TopoPID: 93.77
Enterprise connectionOriginal rules: 75.05Extension: 94.77
OPEN100 connectionOriginal rules: 87.32Extension: 90.86

Identity-aware recognition is evaluated on 524 devices across 45 industrial drawings with a common semantic verifier (Table 2). Connection comparisons hold correctly recognized equipment regions fixed: 489 on the enterprise data and 59 on OPEN100 (Table 5). OPEN100 uses the unchanged enterprise configuration across all 12 drawings.

The recognition comparison supports the value of structure-guided proposals. Component ablations explain how they improve: density alone reaches 75.11% F1; adding visual residuals or refinement raises this to 84.43% or 82.89%, respectively, and their combination gives the strongest result (Table 3). Visual support and refinement thus complement endpoint concentration in resolving equipment extent.

Text/roster controls further distinguish equipment acceptance from engineering identity assignment. On 512 fixed positive candidates, acceptance remains 93.55-98.44% across the four conditions. With image text retained, supplying the roster increases joint name-tag correctness from 6.84% to 95.51%; retaining the roster while masking all characters in both views yields 79.69%. These results show the complementary value of drawing text and drawing-specific semantic knowledge for determining equipment identity (Section 5.2 and Appendix B).

Direct connection evaluation establishes the corresponding role of structural constraints after recognition. Local visual reasoning distinguishes process lines from equipment outlines and ambiguous structures, while explicit constraints govern graph updates. The improved attachment quality on both datasets complements association retrieval by assessing whether individual connections are correct.

3. Manuscript revisions, public evidence, and computational cost.

The revised Introduction, Discussion, and Conclusion connect these findings to the structural-inference contribution. Our anonymous results repository provides OPEN100 recognition and connection outputs, per-drawing summaries, and visualizations, enabling inspection of individual predictions and graph relations.

We also measured token usage, API fees, and elapsed time for recognition and constraint-guided connection. Summing the two stage averages gives 65.1 s and CNY 0.0731 per OPEN100 drawing, and 138.6 s and CNY 0.2688 per enterprise drawing. Elapsed time includes code execution and API waiting time.

We would appreciate your feedback on whether these clarifications and additional results address the main concerns, and would be happy to clarify any remaining questions.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.