Point Cloud Deep Learning Network Based on Local Domain Multi-Level Feature

Round 1

Reviewer 1 Report

A review on

‘Point Cloud Deep Learning Network based on Multi-Level Feature Fusion in Local Domain’ (Ref. no: ID: applsci-2633610)’, a research paper submitted to Applied Sciences (MDPI)

 

The paper titled "Point Cloud Deep Learning Network based on Multi-Level Feature Fusion in Local Domain" presents a compelling exploration into the field of point cloud analysis through the integration of multi-level feature fusion in a local domain. Point cloud data has gained significant attention in recent years, owing to its applicability in various domains such as robotics, autonomous vehicles, and 3D modeling. This paper evaluates the contributions, methodologies, strengths, and areas for improvement in the paper.

 

The paper makes several notable contributions to the field of point cloud analysis:

The methods used by the authors are: Dynamic graph convolutional network and, further, feature points are used as input in reconstructing the local domain to obtain the low-dimensional relationship information between the feature points, of course, it is mapped to higher dimensions using the MLP to further extract the point cloud features. By doing so, the authors claim that the proposed method has better performance in the field of semantic segmentation. The proposed network demonstrates impressive results in capturing localized context information (up to some extent), which is crucial for tasks like object recognition, segmentation, and scene understanding. This localized context awareness is achieved through the fusion of local features, enabling the network to make more informed decisions.

Most importantly, the authors accepted deliberately the shortcomings of the proposed methods, in particularly in extracting the features, if the edge points of some objects are close to each other and they have similar geometric features.    

                Based on the above aspects, I inform that this paper be accepted, but there are a few minor issues that need to be clarified beforehand. I am willing to review this paper once again to check whether the points raised by me are properly dealt or not.  

 

Minor Issues:

1.       Barring one recent paper, I did not find recent papers. I , therefore, request the authors to include two to three 2023 papers in the revised version of this paper

2.        I also recommend the authors to write the organization of the article at an appropriate place

3.       I found an incomplete sentence in page 7, line 248. ‘The dataset with 40 categories and 12311 models.’ Kindly write that sentence in a more meaningful manner.

4.       Page 15: Lines 456-457, a confusion sentence. ‘The man-made terrain and 456 natural terrain are accurately segmented by the proposed method or something else‘?? Kindly re-write.   

5.      I recommend the inclusion of the following papers in the revised manuscript

a)      Deng, C.; Peng, Z.; Chen, Z.; Chen, R. Point Cloud Deep Learning Network Based on Balanced Sampling and Hybrid Pooling. Sensors 2023, 23, 981. https://doi.org/10.3390/s23020981

b)      P. S. Brahmanandam, 2021, Prediction of Atmospheric Particulate Matter (PM2.5) Over Beijing, China using Machine Learning Approaches, INTERNATIONAL JOURNAL OF ENGINEERING RESEARCH & TECHNOLOGY (IJERT) ICRADL – 2021. https://doi.org/10.17577/IJERTCONV9IS05094  

 

 

I found a few minor errors that need to be removed before being accepted for publication. 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

The manuscript proposed a method for 3D point cloud classification and segmentation based on deep learning and local multi-level features of point clouds. It is innovative to some extent that integrating the multi-scale neighborhood analysis, edge convolutional layers, and relational convolutional layers for extracting point cloud features and executing deep learning network, followed by the classification of the point cloud. The experimental results show the advantages and prospects of the proposed method in this manuscript. This article has a clear structure, complete content, and reliable conclusions. Anyway, please consider the following minor comments.

Specific Comments

1.       Suggest modifying the title to align with the method mentioned in the abstract, e.g., Point cloud deep learning network based on local domain multi-level feature. The current title is prone to ambiguity and could be mistakenly interpreted as being specific to local point clouds.

2.       Line 15: The method proposed in this manuscript first extracts the neighborhood and edge features of the point cloud. But what are the deeper level edge features?

3.       Line 17: The method proposed in this manuscript is based on local domain features. However, the global features mentioned here are not explained before. Is this contradictory to the method name?

4.       Briefly introduce the dataset used in the abstract using a sentence, rather than directly referencing the dataset name.

5.       Line 17, pay attention to the correct use of punctuation.

6.       Line 147, a brief description of the proposed network architecture should be given in Figure 1. Additionally, a comprehensive flowchart is suggested to be provided.

7.       This work exhibited the quantitative evaluation of the classification results (Table 1 and Figure 4). Could you please add a qualitative assessment?

8.       Section 4.1 is suggested to be moved to Chapter 3.

9.       In Section 4.2, please provide the sketch maps of an exampled point cloud dataset and its corresponding training samples.

10.     In Table 1, the dataset name should be mentioned in the caption. Additionally, the ‘Input’ column can be removed as the input formats of all methods are the same.

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 3 Report

Deep learning has been widely used in the field of point cloud segmentation, Pointnet or pointnet++ are two famous point cloud deep learning networks. In this paper, authors proposes a  Local domain multi-level feature fusion point cloud deep learning network. Authors use dynamic graph convolutional to obtain local neighborhood features of point clouds. Simultaneously, the relation-shape convolution was used to extract deeper level edge features of point clouds. From the comparison, the fusion of the two methods can improve the accuracy of point cloud analysis.

The paper is interesting and well-written. However, a few things are to be considered sincerely before its publication.  

(1) Figure 1 shows that the input of the network is the point set P= {p1, p2,,…, pn} ⊆RF, N is the number of sampled points, and D is the feature dimension of each point of input. But where is the D in Figure 1.

(2) In Figure 1, the dimensional of data is 3. If the dimensional of point cloud is 6,just like [x,y,z,R,G,B]. Can the data become the input data?

(3) The author's explanation of edge convolution is insufficient. Please added the explanation of Edge Convolution and  describe the edge convolution in detail.  

(4) The square of Figure 9 shows that the trigger points segmented by the PointNet, DGCNN and RS-CNN contain the points of other parts.  Figure 9 or Figure 8?

(5) In Eq.(1), what is the ReLU. Why do you use the ReLU and not other functions.

this paper has some minor english language errors need to be corrected.

Author Response

Please see the attachment.

Author Response File: Author Response.docx