Semantic Segmentation and Roof Reconstruction of Urban Buildings Based on LiDAR Point Clouds

Round 1

Reviewer 1 Report

Dear Authors,

The manuscript under the title “Semantic Segmentation and Reconstruction of Urban Buildings Based on LiDAR Point Clouds” presents a 3D modeling framework consisting of three main works: Proposing the ELFA-RandLA-Net based on RandLA-Net for semantic segmentation of building point clouds, describing automatic identification method of roof surface point cloud and, proposing a roof vertical plane inference method for building topology reconstruction.

The manuscript is well-structured and has interesting content for IJGI readers, however, many issues must be addressed to increase the manuscript's quality. The current state of the manuscript will not be enough for publication in IJGI. My concerns are as follows with the related line numbers of the manuscript:

L2 The title indicates the reconstruction of urban buildings. However, the proposed contributions mainly focus on roof reconstruction. So the title creates a misleading impression. It must be revised.

L21 What is the name of the traditional method?

L25 Authors discuss many references in the Introduction section. However, references are cited as groups such as [4-8], [9-14] and [15-18]. In the Related work, some references were separated and elaborated. But some of them are still ignored for example [6,8,10,14,...]. It needs more explanation.

L74 Satisfactory results from [19-24] must be given quantitatively in a table. And explain why you chose the RandLA-Net. 

L83-87 Authors criticize RANSAC as leading to the problem of competing segmentation planes and ultimately obtaining sub-optimal segmentation results. However, they claim that they used RANSAC for this reason. It does not make sense.

L116 What is the special consideration?

L239 Figure 5 I did not understand the subplot 15d (topology graph). Please elaborate on it.

L328 (A1, B1, C1, D1) are written recursively on the whole line. Is it correct?

L377 There are alternative benchmark datasets about point cloud segmentation. I consider that a second benchmark dataset must be applied to validate the proposed architecture and algorithms.

L377 To reason about the proposed model, the amount, and the content of the training, test, and validation data sets must be explained. 

L386 Authors should expand the metrics such as F1, kappa coef, and precision; because other researchers might need those values to compare their new methods’ performance. Also, the equations of all utilized metrics must be given.

L406 Why don’t you compare your proposed model with the other semantic segmentation architectures and/or models that show good performances such as DGCNN and ConvPoint in the literature?  Please review and consider the state-of-the-art methods to compare with yours.

L407 The smaller value 57.01 was highlighted in the undefined column of Table 1. Please check it.

L 474 Selecting the extreme P2M and MP2 values is a little tricky in the figure. I recommend highlighting max and min values. 

L478 Figure 15 depicts the point cloud and the structured model together. The locations that observed worse performance should be marked in the image.

L484 Future work perspectives should also be discussed in the Conclusions. 

L516 It is tough to evaluate the references because the journal names are not written. I do not expect authors to require a reference format at this stage, but more is needed.

Author Response

Please refer to the attached document for feedback and revised manuscript

Author Response File: Author Response.docx

Reviewer 2 Report

In this paper, the authors propose a semantic segmentation network with enhanced local feature aggregation to locate local details based on geometric and semantic characteristics of urban point clouds that enable a better end-to-end point cloud acquisition quality of building categories. 

The authors studied semantic segmentation techniques based on the aggregation of improved local features to achieve efficient semantic segmentation of building point clouds and propose a method for automatic point cloud identification and a method for reconstructing building topology from the roof surface to extract the roof point cloud plane.

Within this framework, and in order to solve certain problems such as the failure of building roof reconstruction, a method of inferring the vertical plane of the roof to guarantee the accuracy of building topology reconstruction has been proposed. To achieve this goal, the authors have designed a bimodal wireless sensor network that enables remote fire monitoring and warning.

The presentation of the scientific literature on similar or related work is exhaustive but sufficient. The referenced articles are recent and directly related to the work presented in this article. 

This article reflects an undeniable practical dimension, the research methodology is relevant and the results are clear and satisfactory. The text is well written. 

The introduction and conclusion of the article are well synthesized and reflect the essential ideas and methods used to obtain the desired results.

Author Response

Please refer to the attached document for feedback and revised manuscript

Author Response File: Author Response.docx

Reviewer 3 Report

This paper discusses the semantic segmentation method to distinguish buildings from dense point clouds with further clustering method to extract single buildings based on accurate reconstruction of roof planes.

The article is well structured with all necessary sections present and described.

A simple introduction to the topic is given in the paper.

The methodology is well-defined and described.

It clearly explains the source of data and how they were processed.

There is a neat explanation of the authors’ method, both with the use of figures and with text and equations. Without the data, it is hard to judge if the method does what it is supposed to do, however from the description it is an improvement over existing methods.

Everything is effectively summarized.

Possible improvements:

The experimental data have unusual point data density which is very rarely found in the existing data worldwide. It would be interesting to see how well the method works on datasets with fewer points per square meter,

For me, it is hard to distinguish the bush class from the tree class, due to the very close colour within the four green classes. Authors should consider different shades.

Also 'undefined' class would benefit from a more distinctive colour

Some paragraphs are hard to read due to therir technicality with little description or example.

Line 271 - possible editing error

Final thoughts:

IoU almost at 90% for buildings is a very good achievement.

The results section is well written.

The workflow brings the much better accuracy of point cloud classification and semantic segmentation which will be a major issue in the coming years with more and more data available and revisit times thus a need for very good automatic data processing algorithms such as the one described in the paper.

I recommend that this paper be accepted and published.

Author Response

Please refer to the attached document for feedback and revised manuscript

Author Response File: Author Response.docx

Reviewer 4 Report

This paper describes the whole process for modeling 3D buildings from LiDAR point clouds. Starting from the  point cloud of an urban area, the algorithm is able to recognize the individual buildings, identify the roof, obtain the 3D vector representation of the roof and then create the basic 3D model of the building, as well as adding topology to the different structural elements of the building.

The paper is well structured. However, some issues should be solved.

An important problem in the literature is solved, however, the bibliography describes each of the parts of the process instead of paper describing the whole process, from point cloud to 3D vectorial models. 

The paper is difficult to follow in some sections.  Sections 3.2 and 3.3 are not easy to understand. These concepts are supposed to be explained in the papers of the bibliography, however they are not well reflected in the text. For better understanding, the paper needs more images reflecting the processes.  It should be also interesting to show more in detail how the point cloud is being processed along the whole method. 

The paper does not provide information about the ML process. For instance, how do you choose the K value in Figure 3? It is not possible to understand the whole process. 

There are no references in the text to Figures 6.a) and 6.b). Use a), b), etc. instead of numbers  when referring to the figures 7.a), 7,b, etc. Letters in Figure 8 are not recognizable. 

Lines 335-337 reference to nonexistent letters in Figure 7. Equations 15, 16 and 17 appear crowded without additional explanation. Figure 14 should have the same orientation as Figure 13. 

On the other hand, it is supposed that the paper improves the methods described in paper such as [21] and [25], however these improvements have not been described in the text. This makes it difficult to know the contribution of the paper. It seems to be the implementation of several techniques without technical advancements. You should highlight these improvements. 

It should be also noted the case of cantilevered roofs, where the vertical plane of the building is inwards. How does the method work with floating roofs?. Compared to paper [21], the buildings obtained in this paper seem to be more simple. 

Author Response

Please refer to the attached document for feedback and revised manuscript

Author Response File: Author Response.docx

Reviewer 5 Report

This paper presents a whole solution of reconstructing 3D buildings from the airborne LiDAR point cloud. In the first step, author proposes an improved deep learning solution of semantic segmentation to extract the points of building class. then several steps are proposed to create 3D building models. however, there is no obvious siginficant contributation during the steps of detecting roof plane. Author gives enough details of the whole solution. the step of Roof Vertical Plane Inference is proposed to improve the quality of vertical planes and Roof Plane Topology is used to detect the polygons of roof planes. The paper should highlight the importance of these steps. In the experiment section, the comparasion of this solution and PolyFit is presented. This solution  reconstructs 3D models using the points of roof without using walls, but the dataset includes vertical walls. Therefore, the title and abstract should clearly indicate this solution focuses on building reconstration from roof. 

some points need to improve and correct:

323 inner points? give the definition before use terms

344 345 figure number is confused

figure 8 the text color labelling planes is not clear and better to change color, add text to explain the inner and outer points by their colors

add figure to give better explainaiton of texts between 261 and 273, or move equations and details to appendix

the text size of figure 9 is not consistent with others

there is no text to explain the parameters given in RANSAC to segment planes of roof. 

In section 4.3.2 there is no clear explaination of steps for comparative analysis. It is necessary to give details of conducting experiment, use same parameters to detect planes or not?

check and improve some sentences for better understanding

Author Response

Please refer to the attached document for feedback and revised manuscript

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

Dear Authors,

I would like to express my congratulations for the revisions that address all my concerns.

Author Response

See attachment for details

Author Response File: Author Response.docx

Reviewer 4 Report

I will reply to the authors' numbered answers:

(1)

The paper now adds more references associated with building reconstruction, and now at least,  the paper indicates certain improvements regarding current bibliography. 

New text adds some sentences that cannot be included in the current form (for instance, sentence 165 without verb. Is it some kind of title?). Some other sentences seem to make no sense with a period without a capital letter after it (line 174). The sentences start directly with the reference number, which does not seem to be a good writing style for an article. Revise also sentence 188, etc. In summary, English must be revised. 

(2)

The references to these pages and lines seem to be in a different document to the new version of the paper submitted to the system, and it’s not easy to know the new explanation. Anyway, the best place to add new information is not the caption of figures (as in Figure 2 (including reference), 3 and 4). Definitely,  I cannot see more explanations about KNN and MLP.

(3)

Again, there seems to be a mismatch in the number of lines in the new document. Again, the only new explanation is included in figure captions, the text in these sections is the same. I don’t see new images or explanations in Sections 3.2 and 3.3, despite being the most difficult part of the article to understand. This must be clarified.

(4)

Even though the ML process is at least indicated, the basics of the network are not explained, so Figure 3 is not still sufficiently clear

(5)

Ok

(6)

Ok

(7)

Modifications seem to be in pag. 13 in the document I handle

(8)

Ok

(9)

Again, new text seems to make no sense (lines 492). Figure 9 is not referenced nor explained. The layout of new figures appears to be misaligned. You demonstrate that your method works better, but you don’t explain why, that is, what the improvements consist of and how they have been carried out, that is, the improvements in methodology. 

(10)

Please, reflect on this particular case in the document, may be in open problems or future work

In the new text there are sentences and periods that do not make sense. There are sentences that begin with the number of the bibliography that make the sentences unsuitable for an article and sentences without a verb as a title in the body of the text. 

Author Response

See attachment for details

Author Response File: Author Response.docx