Point Cloud Convolution Network Based on Spatial Location Correspondence

Round 1

Reviewer 1 Report (Previous Reviewer 3)

1. The revised version adds a lot of explanations about the disorder and irregularity of 3D point clouds, and it becomes easier to read and understand. The research evidence also confirms that the author's convolutional network framework based on the corresponding point of the spatial position can solve the disorder and irregularity of the point cloud, so it is convincing.

2. For lidar data producers, they focus on the full coverage of point cloud data production and the maximum achievable point cloud density. Since the author has proposed a convolutional network framework based on corresponding points in spatial position, it should be possible to propose the recommended density and basic requirements for the production of data at different scales or different ground objects. This is the recommended addition.

3. The current research results can achieve good prediction accuracy and model convergence. Is there any exception? Considering that the CNN model has a certain degree of gray box characteristics.

4. Since the LiDAR data are mostly produced by integrating multiple point clouds scanned from different angles or different working periods, if these factors are added to the convolution kernel as additional band information, will it help to improve the accuracy of the classified data?

5. At present, the revised version has reached an acceptable level for publication, but it is suggested that the presentation of figures and tables needs to be beautified.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report (New Reviewer)

In this paper, the authors propose to explore point cloud convolution techniques by neural networks, using spatial location correspondence.
After defining the issues and problems of this question, they propose a complete and well-structured state of the art on the subject.
They then highlight the importance of stability according to the order of the points in the cloud. They then outline a framework for implementing these convolutions, and propose on a few public data sets the results of their experiments, in comparison with the state of the art.

I can identify many typo and English errors that affect the quality of the article;. Some examples: "If we want to speed up or improve the accuracy, how to modify its convolution kernel parameter." missing capital letters ("the other is the kernel-based approach"...), non uniform use of space before citations ("Point Transformer [21], Point Transformer[22]"), strange formulations: "and the N-to-M style was adopted for point clouds." The "*" sign in convolutions seems not aligned with the text (it should be lowered). "The k in Formula (3) can represent not only" (text should be on the same line). 4.3: *D*iscussion. "We chose [...] of kernel points" not well formulated. "Disscussion" -> "Discussion". In figure 7: "convulotion" -> "convolution".

From the scientific part, it is not clear to me how the the correspondence problem with shuffled data has been addressed by the network given in figure 7. This figure is not easy to understand and consider together with the past sections. Moreover in this figure, N, Din, Dout are not defined.

In the next sections, which "defined point convolution" has been used? This is not clearly defined and not very straighforward to identify what is the originality of the proposed approach since it miss some details.

In the figure and tables on the experimental results, you should write if the obtained results are in the test set or trained set.

I noticed that there is no confusion matrix to illustrate the quality of the classifications.

For these different reasons, I find it difficult to evaluate the scientific quality of the article.

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report (New Reviewer)

 

It's an interesting study that I think may be of interest to readers. This work has been reviewed based on comments from three previous reviewers, covering all the requirements they made. In the introduction they explicitly present the contributions of the work that are corroborated throughout the work. 

The design of the research is adequate, providing the methods in an appropriate way and presenting the results with an adequate discussion.

However, the main limitation of the work is that it cannot be applied to point clouds obtained by Lidars (without color information), which are sensors widely used in the literature.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report (New Reviewer)

The authors addressed the main part of reviewers' questions, and even if some part of the article are not typical from the classical way to write a scientific article, the resulting document is interesting.

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

The author presented a study, formulating a convolution network based on the irregularity of spatial location of point cloud. The developed technique presented an overall accuracy of 92.7 % in comparison with other 3D shape classification networks. 

General comments: 

The manuscript lucks of structure. Parts of Materials and methods appear in the introduction. The related work appears after the aims of the study. 

In parallel the objectives- aims of the study are not clearly presented. They should be heavily revised, clarifying the aims of the study.

The experimental plan, is missing from materials and methods. Instead, it seems that partly appears in the beginning of results. Moreover, the details of each semantic segmentation that was used should be also stated in M&M.  Please revise accordingly.

In several equations the variables are not explained, making it very difficult to follow the conceptualization. 

The current state of conclusions it looks like an abstract. The conclusions should recap the main findings of the work, providing the most significant results. Please revise. 

I also included some extra comments in the attached pdf.

Comments for author File: Comments.pdf

Reviewer 2 Report

Line 26: You should make the Introduction section concise. It’s too verbose. You should remove some irrelevant and unnecessary contents in the Introduction section.

 

Line 310: How does the Euclidean space distance is determined? What is the justification for the appropriate distance to use?

 

Line 481: There is no Discussion section in this manuscript.  You should discuss the results and research findings of your study. Without this section, your manuscript is incomplete and cannot be published.

Reviewer 3 Report

This study is based on the assumption that the convolution operation remains the same as long as the correspondence remains the same. Based on this assumption, the mathematical properties of convolution in deep learning are discussed. This is an interesting study that should be of great interest to readers. Here are some suggestions that deserve further discussion or revision:

1. This study proposes a convolutional framework that can be directly applied to point clouds, which takes into account the spatial position correspondence. This approach is very good, but basically because there are many ways to generate point cloud data (such as no-load LiDAR, ground LiDAR, vehicle LiDAR, point clouds generated from stereo pairs of photos), the spatial distribution characteristics of point clouds of these data basically exist There are great differences (such as data leakage or the spatial distribution characteristics of obscured parts, point cloud density), so it is recommended to clearly define or explain the possible impact of LiDAR data in the description.

2. This study discusses the location-based correspondence types and the impact of the number of kernel points and generation methods. The results of the study also echo the assumption that the correspondence remains the same, and the convolution operation remains the same. The study also validates the feasibility of convolutional network design through classification and semantic segmentation tasks. It is suggested to add data presentation to illustrate the subsequent operational efficiency and accuracy of point convolution kernels, planar convolution kernels, and stereo convolution kernels. Because accuracy may not be the only consideration in practical applications, sometimes it is necessary to emphasize operational efficiency and stability.

3. Figures in this study may need to be reproduced in a clearer way (the current picture quality is not very good)

4. ResNET itself is characterized by residual operation, which is essentially different from traditional CNN. Will this cause difficulties in the classification of complex data (such as classification of objects with similar characteristics)

5. Whether the classification presentation method is considered to be included in the Loss fuction curve, because it is very important whether the model calculation results converge and achieve acceptable classification accuracy.

6. Because not all lidar data have RGB characteristics, and some only have XYZ coordinates, is this kind of data not suitable for this research process?