Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Dear Authors,

The manuscript addresses a critical intersection of BIM and deep learning, providing a practical, practice-oriented case study. To meet the standards expected of an archival journal article, it will require a clear review protocol, quantitative evidence in the case study, fuller disclosure of the IFC integration layer, and a balanced presentation of the claims made.

Recommendation. Consideration after major revisions will be necessary, including additional methods and transparency, as well as some quantitative assessment.

Methodology

Fitness for purpose. In reviews, a transparent procedure is paramount. Currently, literature and milestone selections are curated but not reproducible.

Detail and Replicability.

Add a literature search protocol that features databases such as Scopus and Web of Science, including query strings, time window, inclusion/exclusion criteria, the screening process, and a PRISMA-style flow diagram.

Define how you mapped papers to the three lenses - provide a coding sheet or taxonomy.

For Bimetria, document dataset sources and sizes, labelling process, class definitions, training/validation splits, key hyperparameters, hardware and inference latency. Provide at least a minimal JSON schema (field names, types, coordinate frames, units) and the IFC mapping rules (e.g., IfcWall/IfcOpeningElement/IfcDoor/IfcWindow; property sets and quantities).

Limitations and bias. There is a potential conflict of interest because two authors are affiliated with the company whose software is presented. Please include a clear statement of this relationship in the Conflicts section and describe the steps taken to mitigate bias in the Methods section, e.g., independent testing, use of external datasets, and pre-registration of metrics.

Alternative/reinforcing strategies.

Review: Instead, consider a structured mapping, such as task × architecture × data type × metric, to replace general narrative passages.

For Bimetria: Provide a small comparative baseline, such as U-Net vs. HR-Net, with/without ESRGAN pre-processing, reporting standard metrics for each task: pixel IoU/Dice for segmentation, mAP for detection, OCR accuracy/character error rate, MAE/RMSE for quantities. This would give substance to claims of stability/latency and "sufficiency" of CNN/OCR.

Results

Consistency with the method: While the Bimetria pipeline is well described, no quantitative results are presented. It is not easy to gauge performance or generalizability without accuracy, robustness, and timing figures across a representative set of drawings.

Clarity and presentation are good. The figures are useful, such as the era timeline and pipeline stages. Perhaps add one summarising table per lens: i) key tasks and dominant architectures; ii) typical datasets and standard metrics; iii) integration patterns with IFC - object classes, relationships, property sets.

Interpretation and generalisation. Claims such as that CNN/RNN are still the preferred models in practice for "stability and low latency" require more evidence. Please either perform measurements, such as median inference time on a typical A4/A1 sheet, with a variance over 10 projects, or soften the claim to indicate that this is an observation from practice.

Analytical gaps: A small ablation study - e.g., HR-Net vs. U-Net on 50 sheets, with/without OCR filtering, and with/without straightening and super-resolution - would significantly strengthen the case study and help readers apply lessons to their own pipelines.

Discussion and conclusions

Alignment with results: Discussion is logical, but many assertions are based on experience rather than data. Present quantitative evidence and clearly indicate where you are drawing reflective rather than empirical conclusions.

Contribution and relevance. The insight of the "intermediate JSON layer" between perception outputs and IFC is valuable. Documenting its schema and providing a small example (e.g., supplementary material) would increase its utility.

Limitations and future work. You acknowledge that no public benchmarks exist. Consider proposing a minimal, open protocol aligned with your three lenses; for example, 100 varied 2D sheets with ground-truth masks and quantities, 10 site videos with annotations, and five time-series datasets with standard splits. Specify baseline metrics. This moves the paper from observation to community guidance.

Coherence. The historical timeline is instructive, but the sections on recent LLM/transformer developments occasionally stray from BIM practice. Either tie each milestone explicitly to a BIM-relevant implication, such as attention for point clouds or graph learning for IFC topology, or trim speculative parts.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The submitted paper presents a new application called Bimetria, which uses a hybrid convolutional artificial neural networks / optical character recognition solution to generate 3D models with estimates based on two-dimensional drawings. The paper discusses the evolution of convolutional and recurrent artificial neural networks in applications for building information modeling (BIM).

 

The submitted paper deals with the relevant topic of building informatics. However, there are major concerns about its content, so some comments for the Authors are provided below.

 

At the beginning, it should be clearly stated that it is difficult to determine the categorization of the submitted article. For the categorization of a review article, a systematic literature review is missing, while for the categorization of an original scientific article, a complete methodological part is missing. After reviewing the content, it seems that the Authors somehow want to propose an original scientific article, which is why the reviewer's suggestions are given in this direction. In any case, the article needs to be supplemented in depth in terms of content, taking into account the comments in full, because in its current form it is not sufficient to propose it for publication.

 

After reviewing the article, it became clear that the Authors actually proposed one of the approaches that generally falls into the field of so-called ''Active BIM'' (in this article it is about AI-supported BIM). Because a systematic literature review was not fully conducted, there is a doubt about the original scientific contribution of the tool Bimetria. Authors are thus kindly encouraged to conduct a systematic and in-depth review of ‘‘original scientific articles’’ in the Scopus and Web of Science databases over the last decade in the field of ''Active BIM'' and to significantly more clearly highlight which characteristics of the tool Bimetria represent the original scientific contribution to the field of ''Active BIM'' that cannot be found anywhere else.

 

The next major comment concerns the fact that the methodological part of the article is practically completely missing. At the moment, it seems as if the article is based only on the applicative part, which is not enough. Therefore, it is kindly suggested that a full methodological chapter be prepared, in which it is necessary to generally present: (i) a graphical representation of the structure of the Bimetria system; (ii) a detailed explanation of the Bimetria system entities and their mutual connections in the text; (iii) a flowchart that generally presents how the Bimetria system captures input data, processes it step by step, and reports output data; (iv) the methods used by the Bimetria system for data processing; (v) the software used for the development of the Bimetria system; and (vi) the software used by the Bimetria system for its operation. It is kindly suggested that the visual representations be supported by an in-depth methodological explanation.

 

The paper states on page 6, line 205: ‘‘This set of input data is standardized…’’ At this point, the Authors are kindly encouraged to cite the associated standards.

 

The paper states on page 7, line 227: ‘‘The model was trained in a multi-task scheme. ’’ In this regard, the Authors are kindly encouraged to reveal in which software environment the model was trained.

 

The paper states on page 8, line 236: ‘‘The third stage of data recognition on projections is an add-on to the model that is the core of the OCR application, which analyzes the image for text data.’’ Here, the Authors are kindly encouraged to disclose which specific ''add-on'' (software) they are referring to.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

This scientific article contributes to the broad discussion on BIM development and its ongoing automation. It has significant scientific potential in the context of academic work, although it seems that some valuable points have been overstated. Below are a few comments that will help expand the article.

Lines 153 - 158: CNN and RNN in a BIM environment are adequately described in this chapter, but the summary in the indicated lines is insufficient and flattened. Please complete.

Lines 195 - 197: A better term for drawing straightening is calibration. Please indicate which methods can be used to perform this in the context of developing a BIM model and applying Bimetria.

Lines: 184-210: In what native environment are operations performed on 2D data?

Section 3.1, 3.2: An improvement to the described research procedure will be the creation of a roadmap from the Bimetrics stages. Furthermore, the roadmap will fit well into the multiple timelines used in the manuscript.

Line 221: What imperfections, give examples in brackets, graphical representation etc.

Line 229: "In a time acceptable to the user" means in what time? Please provide the average unit of time required to create a model for one floor in the applied, simple design of the building.

Part 3.3: There are no graphical illustrations of verification examples here, please add them.

Overall merit of Part 3: Good describe but not very detailed, even though the content is interesting.

Part 4. Discussion: The discussion section presents the authors' voices on the topic of BIM modeling automation, but it's somewhat concise. It lacks discussions with other authors of scientific articles. This section requires expansion.

Part 5. Conclusions: I suggest you list the answers to the questions in the introduction.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Dear Authors,

All my requests were addressed, so the work can be accepted as it is.

Reviewer 2 Report

Comments and Suggestions for Authors

No further comments for Authors.

Reviewer 3 Report

Comments and Suggestions for Authors

I congratulate the authors on the prompt revision of the manuscript. In its current form, the manuscript is satisfactory to me. For future submissions, I recommend more careful consideration of the reviewer’s comments and indicating the line numbers in the response where the corresponding changes have been made. Best regards.