- 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
-
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.
-
The substantial improvements over fixed-VLM baselines, together with encouraging zero-adaptation performance on OPEN100, demonstrate the effectiveness of structured localization.
Critical weaknesses
-
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.
-
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
-
Can the authors provide component-wise ablations and computational-cost comparisons to isolate the contributions of topology density, visual residuals, and candidate refinement?
-
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
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.
| Dataset | Stage | Input tokens | Output tokens | API cost (CNY) | Time (s) |
|---|---|---|---|---|---|
| OPEN100 | Recognition | 13,515 | 1,686 | 0.0324 | 28.9 |
| OPEN100 | Connection | 18,129 | 1,829 | 0.0407 | 36.2 |
| Enterprise | Recognition | 38,258 | 4,121 | 0.0876 | 47.4 |
| Enterprise | Connection | 86,391 | 6,711 | 0.1812 | 91.2 |
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.
| Method | TP / FP / FN | Precision (%) | Recall (%) | F1 (%) | Query accuracy |
|---|---|---|---|---|---|
| Original rules | 148 / 14 / 29 | 91.36 | 83.62 | 87.32 | 268/315 (85.08%) |
| Constraint-guided reasoning | 154 / 8 / 23 | 95.06 | 87.01 | 90.86 | 279/315 (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.
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.
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.
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.