Higher-Order Conditional Random Fields-Based 3D Semantic Labeling of Airborne Laser-Scanning Point Clouds

Round 1

Reviewer 1 Report

The manuscript presents a novel strategy for LiDAR point cloud segmentation based on the conditional random field framework.

The manuscript is well presented and structured. The methodology is well explained. I appreciate  the accurate report of the results and in particular the discussion between the methods compared.

I have few comments:

There are grammatical errors all around. Please, proofread the manuscript.

Line 125: please provide a reference to DBSCAN and the full name.

In figure 1, the component defined as pairwise potential seems not connected (included) to the CRF optimization

Line 173: please replace “After being clustering” with “After being clustered”

Line 189: the term “density reachable” sounds grammatically wrong. It could be “reachable density” but still its meaning is not clear, even though you explain it in the following line. I would suggest (up to the Authors) to find a different term a bit more descriptive. 

Table 4: It is not clear the meaning of the multiple columns under each category

Author Response

Please find the attached document.

Author Response File: Author Response.pdf

Reviewer 2 Report

This paper is well written, and the authors have provided details of their approach. My main question is that how the proposed method can theoretically result in a better outcome compared to other point-based classification methods provided in Table 3.

Please find below my detailed comments for improving this paper.

A thorough Proofreading seems necessary.

Structure of the paper

I recommend the authors to modify the structure of the paper by moving all information regarding data sets as the first sub-heading of methodology.

Date sets

While it is claimed that the scenes are complex, the visual representation of the data sets doesn’t support this claim. How large are the buildings? how close are the buildings? how large is the study extent? How dense are trees and buildings that may cause occlusion in classification? what is the slope of terrain for each scene? The terrain characteristics affect the complexity of a scene that is ignored in this paper.

Abstract

Please report the quantitative comparison of the performance of your method with other state of the art methods.

List state-of-the-art methods in the abstract

As the measure MF1 is not defined yet, it is more appropriate to describe the measure in two or three words.

Introduction

Please include recent publications for review (2019 papers).

In lines 32 to 35, problem statement is not clear and it is ambiguous that who is going to solve the stated problem.

Please include relevant references for lines 37 to 41.

Line 52, what is your method?

Delete sentence “the reader can…for more details” in lines 85 and 86.

Line 90- define “ non-linear parameter objects”

Line 90- not clear

Line 101 and 102- how and why?

Line 125- spell out DBSCAN and explain it. also, add a reference.

Line 125- not clear how you adopt…

Line 128- why your proposed method is better than state-of-the-art?

Line 140- “high level scale” not clear and not defined.

Line 148- “section 1 reviews…” you already have done the review in introduction so it seems wired to say this in the outline of the paper here. Please restructure the introduction by separating the review as the next section and keeping this paragraph at the end of the introduction

Methodology

Please bring all the information regarding data sets as the first sub-section of the methodology.

Line 166- why this algorithm is used for separation of ground and nonground points?

Is the process illustrated in Figure 1 an automated procedure? If yes, have you made the codes available to the public? If yes, refer to it otherwise explain such details.

Line 173-grammatical error: ”after being clustering”

Line 177-“Eps” and “MinPts” are not defined. First use the complete term then demonstrate the relevant abbreviation for each term.

Line 203- grammatical error: more smaller

Line 203 and 204- are the codes available? If yes, give them a reference.

Figure 2- change the colours: green is more appropriate for trees and brown for buildings

Line 208- “mixed labels” do you mean false labels? If yes, modify it to “false labels”.

Provide reference for all equations in the manuscript.

Line 233- how this procedure is done? The results from coarse clusters are manually imported to K-means clustering software or the whole process is automated. If the whole process is automated, explain the programming language, computer specifications and the time of whole process. Also add a reference to tell the a reader how the codes can be accessed.

Please draw a more detailed diagram demonstrating all algorithms and procedures. Figure 1 is not detailed.

Lines 231 and 232- why these steps are required?

Line 235- “Then we check” how? Manually or automatically?

Figure 3- why the result of k-means is good for further process and achieving better outcome than other available methods? Theoretically it seems not possible as there are two many classes from the k-means.

results and discussion are mixed in this paper. Please separate them. Also, add a section for validation of your outcomes.

Figure 5- not clear. The procedure starts from below images to the top ones. This is very confusing. All the arrows need to show a specific step, algorithm or process.

Table 3- add reference for all the methods and their relevant available codes.

Paragraph starting in line 450 needs references for all equations.

Explain Mf1 and F1-scroe. Are the codes available for these validation methods? If yes, provide references to the codes.

Conclusion

Avoid abbreviations in conclusion such as CRF, DBSCAN, Mf1 and F1-scroe.

Author Response

Please find the attached document.

Author Response File: Author Response.pdf

Reviewer 3 Report

In my point of view, it's a good paper that combine steps of hierarchical clustering and high order CRFs labelling for point clouds classification. It could be used as referece in future acheivements.

Author Response

Please find the attached document.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

The authors have addressed most of my comments. I encaurage the authors to consider the following improvements:

Abstract

Please delete the references in the abstract.

Please add your response to point 3 to the paper where I found your method unclear.

Instead of using “adopt” for point 8, please reword your explanation in your reply sth such as “ we segmented  framework  simultaneously  combines  the  existing  classic  algorithms  with  the  proposed  probability  density  clustering  algorithm”

Methodology

Point 2 and 6. The provided link is Chinese language and I could not follow the steps asking to select options!

Please add the diagram in point 11 to the paper and explain it.

Please add the point 14 explanantion to the paper where you want to justify using K-means.

Author Response

Please see the attached document.

Author Response File: Author Response.pdf