Performance Analysis of Point Cloud Edge Detection for Architectural Component Recognition

Round 1

Reviewer 1 Report (Previous Reviewer 1)

Comments and Suggestions for Authors

The paper offers a comprehensive comparison investigation of the efficacy of RANSAC and DBSCAN algorithms in edge extraction from 3D point cloud data, specifically within the context of architectural component analysis.  The subject matter is pertinent and timely, particularly in academic and practical realms, given the increasing implementation of digital twin technologies, laser scanning, and BIM within the design and engineering industries.  The study illustrates a robust and meticulously designed methodological framework, grounded in authentic data obtained by laser scanners and processed using a SOR pre-processing algorithm.  The decision to concentrate the research on eight pillars situated in various interior locations aligns with investigating the algorithms' effectiveness in realistic settings, while ensuring enough control over experimental variables.

The execution of algorithms within a Python framework, the application of defined parameters for both methodologies, and the choice of evaluation metrics, including the quantity of identified edges, average linearity, and correctness concerning density, render the work robust and reproducible.  Implementing quantitative measures derived from PCA analysis to assess the geometric consistency of edge detection and using indicators like the mean distance error (MDE) to define spatial accuracy is praiseworthy.  The behavioural distinctions between RANSAC and DBSCAN are articulated comprehensively and supported by persuasive reasoning.  RANSAC is very effective in organised settings and following denoising processes, but DBSCAN has heightened sensitivity to local density fluctuations and a degree of adaptability in irregular environments; however, it tends to over-detect.

The work contains several significant flaws that, if rectified, might improve its scientific quality and clarity of presentation.  The generalisability of the results seems constrained, given that all data were gathered from a singular architectural feature (interior rectangular pillars).  While this decision is rationalised by the need to uphold controlled experimental settings, it is evident that the results cannot be reliably generalised to other structural components, like curved walls, beams, or intricate ornamental features.  Furthermore, no external reference (ground truth) exists, even partially, to confirm the absolute correctness of the edge detection outcomes.  A basic overlay with a BIM model or established CAD outlines would have offered a valuable foundation for more impartial assessments, particularly given the lack of manually annotated data.

The lack of a sensitivity analysis for the critical parameters of the algorithms is seen, especially for DBSCAN, which is particularly sensitive to the selection of eps and MinPts values.  The offered values lack a detailed explanation for their choice, and there is no demonstration of how performance varies among them.  A quick parametric analysis or an example illustrating the consequences of a bad decision would have been beneficial for this purpose.

Another aspect that warrants consideration is the presentation of the results.  Despite the abundance and organisation of the tables, their numbering does not align with the order in which they appear in the text.  The numbering of tables 3 and 4, found in the sections on noise robustness, is repeated and conflicts with that employed in the subsequent density analysis.  The numbering mistake requires meticulous correction, since it impedes the reader's ability to follow the logical progression of the parts and may lead to misunderstanding in data interpretation.  A comprehensive evaluation of the table numbering is advised to ensure alignment with the content and the sequence of display.

Stylistically, the manuscript's language is predominantly straightforward and technical; nonetheless, many paragraphs exhibit redundancy, especially in the findings section, and may benefit from simplification or condensation.  Numerous remarks articulated in the tables are repeated discursively without offering extra interpretive value.  A more comprehensive synthesis at these junctures would enhance the text's fluidity and immediacy.  Furthermore, the article would be improved by including a figure that visually represents examples of the results achieved (for instance, detected edges superimposed on the point cloud), even in symbolic or schematic form, to assist the reader in visualising the distinctions between the algorithms.

The article is a compelling and substantiated addition to point cloud analysis in architectural applications.  The methodology selection is suitable, the analyses are comprehensive, and the reasoning aligns with the articulated aims.  The paper merits publishing; nonetheless, it needs modest adjustments to rectify mistakes in table numbering, enhance clarity in some sections, and fortify the critical discourse on the methodological limits of the study.  Should these elements be sufficiently handled, the paper may be a valuable resource for scholars and professionals engaged in edge detection within intricate three-dimensional settings. 

Author Response

Thank you very much for taking the time to review this manuscript.

We carefully reviewed each comment and revised the manuscript accordingly. 

Please see the attachment for details.

Author Response File: Author Response.pdf

Reviewer 2 Report (New Reviewer)

Comments and Suggestions for Authors

Kim &Yun, Performance Analysis of Point Cloud Edge Detection for Architectural Component Recognition

 

MDPI, Applied Sciences

 

1. General Observations

 

1. What is the main question addressed by the research?

The objective of this study is to select representative edge detection algorithms. This study conducted a comparative analysis of two edge detection algorithms—RANSAC and DBSCAN—using 3D point cloud data acquired from eight interior pillars of a building. The evaluation was based on three performance criteria: edge detection quality, noise robustness, and density-based performance. The authors comprehensively analyze the performance of two algorithms to propose an optimized method for accurately recognizing specific architectural components.

 

 

2. What parts do you consider original or relevant to the field? What specific gap in the field does the paper address?

The use and application of digital twin technologies, 3D scanning and point cloud data in the construction industry is welcome and original. Edge detection is an essential process for effectively analyzing 3D point cloud data. However, unlike in traditional 2D environments, the irregularity and noise inherent in 3D data necessitate the use of robust algorithms. Accordingly, this study applies two algorithms: the Random Sample Consensus (RANSAC) algorithm, which is commonly used for geometry-based edge detection, and the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm, which is utilized for density-based edge detection. The RANSAC algorithm demonstrates high accuracy when applied to geometrically well-defined structures such as lines or planes, enabling effective edge extraction through precise linear fitting. In contrast, the DBSCAN algorithm identifies edges based on data density, allowing it to reflect the spatial distribution of points without assuming explicit geometric models, and showing strong applicability even in noisy environments. 

 

 

3. What does it add to the subject area compared with other published material?

The authors’ approach is basically opening a whole new world of applications. Duanmu et al. aimed to overcome the limitations of conventional PLS (Personal Laser Scanning) by introducing a new preprocessing algorithm called ANPDA (Annular Neighboring Points Distribution Analysis), which improved edge detection accuracy. Pan et al. proposed a method that utilized graph convolution and edge-aware features to reconstruct point cloud data. The approach was computationally expensive and susceptible to misinterpreting noise or erroneous measurements as actual edges. Han et al. proposed a method that defines neighborhood relationships using an octree structure and evaluates the significance of each point based on its normal vector. This approach incrementally removed non-edge points by leveraging geometric information. The study achieved an improved edge retention compared to prior techniques but also presented a drawback of increased computational time.

 

 

4. What specific improvements should the authors consider regarding the methodology?

Geometry-based and density-based methods employ fundamentally different strategies, leading to differences in edge detection accuracy and data preservation performance. Based on a comprehensive review of existing research, the RANSAC algorithm—founded on geometric characteristics—and the DBSCAN algorithm—which utilizes point density—were determined to be suitable for edge detection of architectural elements. The authors combined two complementary techniques to solve the problem. The paper is sufficiently broad that I cannot recommend any additional improvements.

 

 

5. Are the conclusions consistent with the evidence and arguments presented?

Conclusions are logical summary of the presented developments. Advantages and disadvantages of both methods (the RANSAC algorithm—founded on geometric characteristics—and the DBSCAN algorithm—which utilizes point density) are given, including summaries in tables. Cases when one method is advantageous to the other are described.

 

 

6. Were all the main questions posed addressed? By which specific experiments?

The paper answers all the questions that it proposes as solution. The authors set out to analyze the architectural object using edges. Their speed and accuracy are better that the current techniques. Duanmu et al. aimed to overcome the limitations of conventional PLS (Personal Laser Scanning) by introducing a new preprocessing algorithm called ANPDA (Annular Neighboring Points Distribution Analysis), which improved edge detection accuracy. Pan et al. proposed a method that utilized graph convolution and edge-aware features to reconstruct point cloud data. The approach was computationally expensive and susceptible to misinterpreting noise or erroneous measurements as actual edges. Han et al. proposed a method that defines neighborhood relationships using an octree structure and evaluates the significance of each point based on its normal vector. This approach incrementally removed non-edge points by leveraging geometric information. The study achieved an improved edge retention compared to prior techniques but also presented a drawback of increased computational time.

 

 

7. Are the references appropriate?

Yes.

 

 

8. Comments on the tables and figures and the quality of the data?

Figure 1, 2. This reader does not like a black background. The green lettering is illegible.

Figure 2. The last set of blocks is illegible.

All figures. Figure caption should not have each word capitalized.

All figures, including Figure 5. Quantity shown in vertical axes should be indicated with units.

Figure 5 and Figure 5. Two figures re denoted fig. 5.

Second Figure 5. Vertical axes should be marked, and the light green needs to be darker.

 

Tables are fine.

 

 

9. Flaws in methods.

The methods are very carefully designed, incorporating

 

 

10. Comparison with existing techniques.

The paper represents a though comparison and concludes what methods seem to be most appropriate. Duanmu et al. aimed to overcome the limitations of conventional PLS (Personal Laser Scanning) by introducing a new preprocessing algorithm called ANPDA (Annular Neighboring Points Distribution Analysis), which improved edge detection accuracy. Pan et al. proposed a method that utilized graph convolution and edge-aware features to reconstruct point cloud data. The approach was computationally expensive and susceptible to misinterpreting noise or erroneous measurements as actual edges. Han et al. proposed a method that defines neighborhood relationships using an octree structure and evaluates the significance of each point based on its normal vector. This approach incrementally removed non-edge points by leveraging geometric information. The study achieved an improved edge retention compared to prior techniques but also presented a drawback of increased computational time. The existing methos are incorporated inti the paper.

 

 

11. Conclusion and Limitations

 This section should be broken in two separate sections, Discussion and Conclusion.

 

 

12. Sec. 6. Patents

Incorrect title.

 

13. English

Throughout the paper: this reader would rather see »immunity to noise«,  rather than »noise immunity«.

 

 

Recommendation

Publish with minimal changes at authors' discretion.

 

Author Response

Thank you very much for taking the time to review this manuscript.

We carefully reviewed each comment and revised the manuscript accordingly. 

Please see the attachment for details.

Author Response File: Author Response.pdf

Reviewer 3 Report (New Reviewer)

Comments and Suggestions for Authors

The structure of the paper is appropriate; therefore, I will present my comments and suggestions in reference to the relevant sections and parts of the manuscript:

Introduction – (in my opinion, the weakest part of the paper)

a1. The first paragraph of the introduction does not fit the scope of the article. By invoking the broad concept of the "digital twin," the authors should define what it is and in what context it is or will be applied. However, this article is not focused on that subject. The concept may be mentioned as a direction for future research — for example, in the conclusion — but here, it may mislead the reader into thinking that the study deals with this topic, which it does not.

a2. More attention should be given to 3D models — their general definition, modeling methods, types of models, and levels of detail. The concept of an "edge" in a 3D model context should be explained, including how to generalize it efficiently and without losing information, optimize its course, and how it is extracted from different data sources using various tools, techniques, and algorithms. There is a wealth of approaches in this field, but none of these aspects are addressed in the introduction.

a3. The paragraph covering lines 43–50 is overly generalized. It outlines applications of algorithms, point clouds, and edge detection in 3D modeling in such broad terms that no specific information can be extracted. It would be more useful to clearly state the research problem being addressed or identify the gap in the literature. While the current version gestures toward such intentions, it lacks specifics.

4 In lines 52–53, the authors state that the goal of the research is to “propose an optimized method for accurately recognizing specific architectural components” — which is a reasonable objective, but it lacks clarification: what is the intended level of accuracy? What scope? Which LOD (Level of Detail) is being targeted?

5 Section 1.2 is not based on a literature review, even though there is a significant body of work available on this topic. It only provides generalized information, supplemented by assumptions that should be supported by references to relevant literature.

a6. On what basis were these two algorithms selected? What other algorithms were considered? What research was conducted to justify this choice? This needs to be supported by a proper review of the literature, previous studies, and industry insights.

 

B, Related Studies – a highly chaotic and thematically disorganized section

b1. The section begins with advanced algorithms described in isolated publications, while neglecting fundamental methods, their interpretations, and scope of use. These basic methods are often the foundation (through minor improvements) of the more sophisticated solutions discussed later. Why are there no references to core algorithms like RANSAC (https://doi.org/10.1145/358669.358692), LSF, Region Growing, normal vector analysis, or the Hough Transform? A review of these would answer many of the research questions posed.

b2. I suggest restructuring this section: start with the basics, then move to their extensions, and finally present advanced solutions and their applications. The goal of a literature review is to identify what is already known and where the gaps or research niches lie.

 

Research Methodology

c1. The description of the RANSAC algorithm (lines 278–290) should be supported by references to relevant literature.

c2. The same applies to the description of DBSCAN (lines 313–323).

 

Research Results and Analysis

d1. Why were graphical results of the edge detection not presented?

2 Where is the comparison against a reference base — the actual dimensions and shape of the object? One cannot discuss the effectiveness of an algorithm without validation and verification of the results.

d3. Incorrect table numbering: 3, 4, 5, 6 — this should be corrected.

 

ee Section 4.1

E1. A measurement noise level of 1–2% is standard for terrestrial laser scanning (TLS). In laboratory conditions, such noise should not significantly affect the results — especially in the case of RANSAC, which is known for its robustness to noise (I recommend citing relevant literature). Therefore, the result was fairly predictable — but should still be discussed in relation to findings from other studies.

 

F . Section 4.2

f1. Reducing point cloud density through percentage-based thinning is not the best approach. This method randomly eliminates points — even in already sparse or incomplete areas of the cloud. It is only valid when point density is uniform across the dataset, which has not been confirmed here (unless neighborhood analysis was conducted?). I recommend instead using distance-based downsampling (e.g., every 1 mm, 1 cm, or 1 dm), which would ensure uniform point density throughout the cloud.

 

Discussion

g1. Where is the discussion of the obtained results in the context of other studies and existing literature? This is a crucial part of any scientific article — it’s not enough to simply report results; they must be interpreted and compared.

 

H. Why is there a “Patents” section? Does it relate to the content of this article?

 

Suggestions for Further Research:

Z1: The experiment is limited in scale: 8 similar objects (columns). There is no object diversity, no comparative sample, and no validation on other geometries. This is insufficient for a reliable statistical analysis. Merely comparing the number of detected edges is not enough — one must also analyze their accuracy, positioning, reliability, and repeatability.

Z2: The article lacks comparison with more advanced methods, such as machine learning or deep learning approaches — even in a limited scope. The literature contains numerous modern methods for edge detection in point clouds. Comparing only two classical methods provides an incomplete perspective.

Z3: I encourage a deeper review of the literature. This will aid not only in drafting the introduction and state-of-the-art review but — most importantly — in forming a discussion of the results.

Z4: When evaluating algorithms, it is essential to present the used parameters, implementation details, source code (or at least the software used), and — above all — to show the results, not merely describe what was done.

 

I wish the authors success in their future research ! The topic is undoubtedly interesting, but in its current form, this manuscript is not yet suitable as a complete scientific publication. The work requires significant supplementation, in-depth analysis, and refinement.

Author Response

Thank you very much for taking the time to review this manuscript.

We carefully reviewed each comment and revised the manuscript accordingly. 

Please see the attachment for details.

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report (New Reviewer)

Comments and Suggestions for Authors

Thank you to the Authors for implementing the revisions, partially addressing the earlier comments, and providing responses. The article looks significantly better than in the previous version; however, it still contains substantial substantive gaps and errors that require correction.

a) Introduction

a1. The introduction includes examples from the literature and hybrid RANSAC–DBSCAN solutions; however, there is still a lack of a systematic review of the fundamental families of methods (RANSAC, LSF, region growing, Hough transform, normal curvature analysis) with reference to their application in architectural indoor conditions.

a2. It is worth explicitly linking the works cited in the review with the metrics used in the article (linearity, MDE, edge detection accuracy), explaining why the selected indicators accurately describe the quality of architectural geometry representation.

 

b) Research Methodology

b1. The BLK360 scanner used has a declared accuracy of 4 mm/10 m. However, in Table 3 the mean C2C deviations relative to the BIM range from 0.11–0.38 m. These are values several dozen times greater than the device’s noise. Units should be added to the table, and possible causes should be discussed (registration errors, BIM inaccuracies, fitting parameters). It would be valuable to provide the registration RMS and ICP fitting quality.

b2. Providing the parameter values for RANSAC and DBSCAN is a plus, but the “preliminary sensitivity test” is not shown. A simple plot or table should be added (e.g., number of edges/linearity as a function of eps, thr, MinPts) to show the stability of results.

b3. There are no formal formulas for:

- Average linearity (how it is calculated from PCA eigenvalues),

- Edge detection accuracy (%) (what denominator, how correct edges are defined),

- MDE (difference relative to C2C).

b4. The abstract refers to a “reproducible benchmark,” but the article states “Data Availability: Not applicable.” This is contradictory.

b5. In Section 3.4.2 the BIM is presented as a “reliable reference point” reflecting actual geometry. However, in “Limitations” it is acknowledged that it was created from CAD drawings, not from precise as-built measurements. This is an internal contradiction.

 

c) Results and Analysis

c1. Units should be added and it should be explained why the values in Table 3 are so large compared to the declared accuracy of the scanner.

C2. Edge detection accuracy (%) — without a clear definition and matching rule to the reference (e.g., number of expected edges, angular/positional tolerance), it is difficult to assess what the table values mean.

C3. In the tables concerning noise robustness (DBSCAN) there are likely erroneous values. Please check their correctness.

C4,. Section 4.1 (noise robustness) — only 1–3% of points were removed using SOR, and conclusions are based on minimal differences. Confidence intervals and a simple statistical test should be added, the difference between this experiment and the density analysis (where percentage thinning was mentioned) should be clarified, and literature on RANSAC noise robustness should be referenced.

c5 Section 4.2 (density analysis) — please provide the actual density values in each group (high/low) and justify the choice of the 20% threshold.

 

d) Results

d1. A more complete graphical presentation of results is still lacking.

 

e) Discussion

e1. Compare the obtained values (C2C, linearity) with typical ones from the literature and discuss why the mean C2C values are so high.

e2. A BIM created from CAD drawings is not ground truth; ideally, at least part of the cases should be verified by independent measurement.

Author Response

We sincerely appreciate the reviewer’s thorough and meticulous feedback, and we have made every effort to revise the manuscript in line with the suggestions provided.
Please refer to the attached file for detailed responses and corresponding revisions.

Author Response File: Author Response.pdf

Round 3

Reviewer 3 Report (New Reviewer)

Comments and Suggestions for Authors

Thank you to the Authors for the revisions and responses; the article looks markedly better, but it still contains significant shortcomings and requires substantial corrections:

a) Introduction
a1. A systematic map of concepts and clear definitions of the metrics are missing at the outset, so readers can correctly interpret subsequent results and conclusions.
a2) A clear justification for the choice of algorithms relative to other method families is missing, since tool selection should follow explicit criteria and assumptions for indoor data.

b) Related Studies
In my view, a review of foundational methods (RANSAC/LSF/region growing/Hough/normals) with their indoor application scope is missing, as this situates the study in context and identifies the real research gap.

c) Methodology
c1. Treating BIM as ground truth is inappropriate, because an as-designed model does not guarantee as-built conformity and requires independent validation or a manually prepared reference set.
c2. Definitions of the metrics and thresholds are incomplete, as they should include formulas, units, and matching rules to enable reproducibility and comparability.
c3. Registration/ICP quality is insufficiently documented, because interpreting deviations without fit information remains ambiguous.

d) Results and Analysis
The interpretation of the metrics is ambiguous, because it is not clearly defined what counts as a “correct edge” and within what tolerance.

e) Noise Robustness
The procedure description is unclear, because the “noise removal” and “density” processes are intermingled, which hinders interpretation; please clarify the description.

f) Validation against BIM
Units and tolerances are not fully harmonized between the text and the tables; some tables lack consistent column descriptions and comparison conditions.

g) Discussion
There is no synthesis of results with the literature and no discussion of discrepancies; the lack of a broader comparison prevents assessing whether the observations reported by the Authors are typical or driven by specific settings and data.

h) Conclusions / Limitations
The conclusions are too generic, as they should clearly present not only strengths but also the limitations of the metrics, procedures, and experimental scope.

 

Author Response

We sincerely appreciate the reviewer’s thorough and meticulous feedback, and we have made every effort to revise the manuscript in line with the suggestions provided.

Please refer to the attached file for detailed responses and corresponding revisions.

Author Response File: Author Response.pdf

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The article "Performance Analysis of Point Cloud Edge Detection for Architectural Component Recognition" looks at a big problem in point cloud analysis for finding architectural parts by comparing and contrasting two methods, RANSAC and DBSCAN. The work is well-organised and articulates a clear technique in the division of experimental phases; yet, several significant difficulties arise that undermine its scientific validity and relevance to the current literature.
A primary issue is the methodological clarity and repeatability of the experiment. The explanation of the data preprocessing phase is inadequate, hindering other researchers' ability to duplicate the results. No comprehensive information is offered regarding the settings used for capturing and refining the point clouds, nor are there any challenges faced during data acquisition and processing. Also, the dataset only includes data from a single small indoor area (60 m²), which makes it hard to say how well the results will apply to more complex and varied situations.
A significant difficulty pertains to the choice of evaluation metrics. Even though important factors like accuracy, processing speed, and noise robustness were taken into account, there is no theoretical reason for choosing these values. There isn't enough information given about Intersection over Union (IoU) and Mean Distance Error (MDE), and it's still not clear if these measurements can be used to accurately check the quality of edge detection in architectural point clouds. Moreover, the introduction of fake noise to assess the algorithms' robustness lacks verification of its alignment with the actual noise present in the LiDAR-acquired data, thus undermining the evaluation's validity.

The comparison between the two algorithms is limited to a singular structural type, specifically architectural columns, but the conclusions of the study also pertain to other structural elements, such as beams and walls, which were not analysed. This oversight constitutes a limitation of the study, as it hinders the ability to draw generalised conclusions regarding the efficacy of the two techniques in identifying various architectural components. An investigation of various architectural features would have been beneficial to evaluate if the performance of the two methods fluctuates based on the geometry and complexity of the buildings.
The analysis of the results exhibits certain deficiencies, since some assertions lack sufficient empirical basis or fail to reference prior studies. There isn't much of a difference between the two algorithms' execution times (2.89s for RANSAC vs. 4.61s for DBSCAN), but this doesn't mean anything because it doesn't take into account external factors like hardware requirements. Moreover, there is no comparison with prior studies that have assessed the computing speed of these algorithms on analogous datasets. This results in another constraint of the study: the absence of current and pertinent scientific references. Despite referencing several studies, the proposed work's unique contribution to the existing literature and its differentiation from others remain unclear. In the absence of a definitive comparison with prior research, assessing the originality of the suggested method and comprehending the actual advancements relative to previously published studies is challenging.
Another issue is the lack of comparison with other techniques for detecting edges in point clouds beyond RANSAC and DBSCAN. Contemporary methodologies, such as neural network-based techniques or hybrid approaches, might have been used as a comparative foundation to more effectively assess the limitations and merits of the two algorithms studied. Deep learning models like SuperPoint and LoFTR are extensively utilised for edge extraction in point clouds, offering enhanced accuracy and noise resilience. The lack of discourse on these alternative techniques diminishes the scientific significance of the study and renders it less pertinent in the contemporary research landscape of this topic.
In conclusion, the work addresses a topic of undeniable significance, but it possesses significant deficiencies that undermine its scientific rigour and its relevance to the current literature. The authors should make the comparisons with other studies stronger, give a better explanation for their choice of evaluation metrics, test the algorithms on a wider range of architectural structures, and think about comparisons with other methods in order to make the manuscript more publishable. A comprehensive rewriting of the manuscript, incorporating enhanced scientific references and a more critical examination of the findings, might substantially elevate the article's quality and establish it as a more credible addition to the domain of point cloud analysis in architectural applications.

Author Response

Thank you very much for taking the time to review this manuscript. In the revised version, all changes in response to the reviewers’ comments have been clearly indicated in red font within the manuscript. We sincerely appreciate the reviewers' insightful feedback, which greatly contributed to improving the overall quality and clarity of the paper.

To enhance readability, we have grouped together comments with similar meanings and addressed them in an integrated manner within the response letter. We hope this format helps convey our revisions more clearly.

For further details, please refer to the attached file.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

COMMENTS AND OPINION:

• It is necessary to standardize the font height in the abstract (lines 12 – 17)!
• Please reduce number of words in Abstract section! According to Template file, Abstract needs to have 200 words maximum!
• Point is missing (line 170)!
• The paper mentions the names of two 3D scanners, LiDAR and LEICA BLK360. It also states that both scanners were used! Was both LiDAR oand LEICA BLK360 used in data collection? If both scanners were used, then it is not clear enough in the paper to explain which data set was collected using one scanner and which data set was collected using the other scanner. The accuracy of both scanners is not the same, therefore the accuracy of the collected data (point clouds) is not the same!
• Figures 2 and 3 are missing (line 247)??!!
• In this research lacks a mathematical model, a mathematical description of the methods that were used:

DBSCAN algorithm model

RANSAC algorithm model

SOR algorithm model

Intersection over Union

Mean Distance Error

Is possible to present a mathematical model for this methods?

• Processing speed often depends on the technical characteristics of the computer (CPU, RAM,...) that runs data processing applications such as numerical analysis applications (FEM-mesh size, complexity of geometry, boundary conditions, sub-modeling) or CAD modeling! Therefore, this part of the data remains unclear?!
• It would be good to conduct research if not only linear characteristics were analyzed, but also non-linear characteristics that would include for example arcs and other geometrical complex forms!
• Figure 4: Please arrange figure structure according to template file! Please see Template file!
• Explain in the paper the method of the artificial noise addition!
• In Table 5 is possible to see different values for DBSCAN algorithm, 39 and 51 detected edges! The same is and for edge detection rate! In text lines 474 and 475 is not the same data! Please reconcile the data!
• In line 514 - Table 6 name needs to be above the table!

Number of references is very weak (11 references)! Is it possible to expand?!

Final opinion:

Please fix the paper according to comments and questions added in this review form!

Comments for author File: Comments.pdf

Author Response

Thank you very much for taking the time to review this manuscript. In the revised version, all changes in response to the reviewers’ comments have been clearly indicated in red font within the manuscript. We sincerely appreciate the reviewers' insightful feedback, which greatly contributed to improving the overall quality and clarity of the paper.

To enhance readability, we have grouped together comments with similar meanings and addressed them in an integrated manner within the response letter. We hope this format helps convey our revisions more clearly.

For further details, please refer to the attached file.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

I have thoroughly reviewed the amended paper and acknowledge with gratitude the modifications implemented in response to prior feedback. The technique section has been notably enhanced by the inclusion of comprehensive details on the data acquisition settings and the pre-processing procedures, rendering the work clearer and more replicable. The elucidation of the assessment metrics has been enhanced, sufficiently rationalising the selection of indicators like mean linearity and the lack of ground truth for the use of traditional metrics, such as the IoU. The paper accurately acknowledges the constraints associated with the dataset's limited size and presents a compelling strategy for expanding the study into more intricate scenarios in the future. The paper now demonstrates an enhanced understanding of the contemporary scientific scene through the introduction of a section that briefly discusses deep learning-based techniques, emphasising their promise while staying beyond the immediate subject of the study.

Nonetheless, while recognising the advancements achieved, I believe certain elements remain susceptible to enhancement. It is necessary to make it clear what the work's original contribution is in relation to what's already been written. The paper has a good amount of technical information, but it's not clear what makes it different from other research. A brief concluding paragraph succinctly emphasising the significance of this study—such as the selection of measurements, the architectural setting employed, or the comparison methodology—would enhance the contribution's impact. Also, the way the text is organised could be better with better synthesis; some sections have unnecessary repetition of numerical data in both the text and the tables, which makes reading less fluid. A summary table or figure juxtaposing the strengths and drawbacks of RANSAC and DBSCAN would provide a visual encapsulation of the results, enhancing the immediacy and use of the study for the reader.
In conclusion, I assert that the work, in its present state, has made significant progress and offers a robust basis. With slight modifications in style and an increased focus on added value rather than cutting-edge innovation, it may be deemed suitable for publishing.

Author Response

Thank you very much for taking the time to review this manuscript. In the revised version, all changes in response to the reviewers’ comments have been clearly indicated in red font within the manuscript. We sincerely appreciate the reviewers' insightful feedback, which greatly contributed to improving the overall quality and clarity of the paper.

To enhance readability, we have grouped together comments with similar meanings and addressed them in an integrated manner within the response letter. We hope this format helps convey our revisions more clearly.

For further details, please refer to the attached file.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The authors significantly improved their research and to a large extent clarified certain confusions. The research presented in this paper is original and I suggest accepting the paper!

I thank the authors for their efforts!

Author Response

We sincerely thank the reviewer for the thoughtful and encouraging feedback. We greatly appreciate your recognition of the originality of our research and your constructive comments, which have been invaluable in improving the clarity and overall quality of the manuscript. Your recommendation to accept the paper is truly encouraging, and we are grateful for your time and effort in reviewing our work.

Thanks, again.