A Building Point Cloud Extraction Algorithm in Complex Scenes

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper proposes an algorithm for extraction of buildings from Lidar data, that is just from the position of the points.

Introduction is sufficient and clear for the reader to get with the problematics on which solution in this article is proposed.

The proposed method is explained in details.

The experimental results are nicely and clearly presented.

Conclusions follow experimental results.

I have a few minor comments and questions:

1. Since proposed method gives better results than two methods (PointNet and PointNet++) and is comparable with HDL-JME-GGO, are there any advantages of the proposed method in regards HDL-JME-GGO method (maybe processing time)?

2. There is word "façade" in a few places instead of "facade".

3. Line 62: DSM is mentioned for the first time, but in the lines 64/65 mentioned achronim is explained.

4. In lines 233/234 it is stated that points are connected counterclockwise, but the arrows are clockwise.

Best regards and good luck.

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Title and keywords:

Vegetated Scenes: mean a scene that contains only vegetation. You want to say that the scene contains a mixture of vegetation and buildings, which is why you missselect the good expression, please replace it. Moreover, the Keywords should be famous words, please remove “Vegetated Scenes” and “mask polygons”

You say: “point set” it must be “point cloud”

 Abstract

Line 16: “mask polygons”: this term is unknown, even if you can’t find it on Google. The use of this term in the abstract makes the abstract vague. Please replace it in the abstract because the abstract should be clear and understandable.

Please avoid using (we, our, and us), use the passive voice. Please check the paper.

The abstract is badly written, where you spend two lines to explain opaquely the suggested approach and then spend all the rest on the introduction result evaluation and comparison. 70% of the abstract should be spent on the approach summary. Please re-write the abstract.

 

 Introduction

You say: “Building point set extraction methods can be classified into two categories based on data sources: single-source methods and multi-source methods.” This classification is correct but not accurate because a single source can be any approach that uses single input data such as aerial images. You want to say “ two family approaches: the first uses exclusively LiDAR data and the second uses

 Auxiliary data in addition to LiDAR point cloud.

I observed that you try to create your terminology lexicon. That is unacceptable when you write a paper. It reflects a weak bibliography research. I am sorry to say that. This issue may appear when you write in a non-English language and then translate the paper into English.   

You say: “However, when dealing with tree points closely attached to buildings, there is a possibility of misclassifying them as buildings.” There are two similar papers to your research that should cited and compared to your work which are:

Tarsha Kurdi, F., Gharineiat, Z., Campbell, G., Awrangjeb, M., Dey, E.K. 2022. Automatic filtering of LiDAR building point cloud in case of trees associated to building roof, Remote Sens. 2022, 14, 430, https://doi.org/10.3390/rs14020430.

Martin-Jimenez, J.; Del Pozo, S.; Sanchez-Aparicio, M.; Laguela, S. Multi-scale roof characterization from Lidar data and aerial orthoimagery: Automatic computation of building photovoltaic capacity. J. Autom. Constr. 2020, 109, 102965.

This paper confirms your paper's motivations.

Line 71 to 79: you cited an approach of building modeling using the RANSAC algorithm which is far from the context of building detection. Please remove it.

Line 80: you suddenly start talking about deep learning algorithms. You should say: “Classification approaches can be also classified into rule-based approaches and machine learning approaches. All last cited approaches belong to rule-based approaches.” Then you can start talking about the machine learning approach.

Furthermore, why do you talk specifically about deep learning, and you neglect all machine learning family?

I disagree with you about your proof that “they require high-quality training data and are sensitive”, because you tested your approach on a few point cloud samples. There is a great probability that your approach provides weak results with other samples. You can say simply that the presence of a machine learning classification algorithm does not represent a reason to stop the research in rule-based approaches.

Lines 103 to 111 should be moved to the method section.

The first point of “main contributions” should be removed because it does not represent a contribution.

 Methods

Line 131: cloth simulation filtering (CSF) algorithm: please cite a reference.

Listen please, the main issue of LiDAR data classification is the huge data volume. You use old and small ISPRS data samples to prove your approach. The new LiDAR data point density is about hundreds of points per square meter. If you develop a new classification strategy that provides good results with 50 million or 100 million points, it could be good, but if you test your approach with a point cloud of one million points, that is outdated (I am sorry to say that).

Figure 2 contains mistakes, where there is a line without an arrow, the term “mask polygon” is not correct, and you should replace it. The image “Ground points in dark yellow and non-ground points in grey” could not give “Coarse extraction of the building point set” by using region growing. The caption of the last image neglects the green, and the purple does not appear in the image.

Line 164: what is the value of G.

Line 162: what do you mean by a particle (do you mean a LiDAR point?). How do you calculate X(t)?

How “b” value can be determined?

Pk is position of adjacent particles, how it can be calculated?

Figure 3: Please add the DSM image beside the shown result to see if it is correct or not.

Line 181: how the curvature is calculated?

How did you determine the threshold values employed in the region growing algorithm?

How did you get the ground truth presented in Figure 5?

In the Alfa shape algorithm, how do you determine the circle radius value?

Please add a section to discuss the long list of used thresholds.

Conclusion

Please discuss the impact of using a great number of thresholds on the suggested method, where these thresholds can take different values from one point cloud to another. How this issue can be solved?

Please discuss all limitations of the suggested approach.

 

Please discuss future works.

Comments on the Quality of English Language

Minor editing of the English language may be required.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

The paper looks much better.

Comments on the Quality of English Language

 Minor editing of the English language may be required.