PAN: Improved PointNet++ for Pavement Crack Information Extraction

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

PAN: Improved PointNet++ for Pavement Crack Information Extraction.

 

The submitted paper was written very well and followed all the steps of the scientific research paper process. I have some minor comments to improve the paper's quality.

 

1. The abstract is too long and some included information part can be moved to introduction section.

2. Authors need to provide detailed information about the training and testing process, and how to solve overfitting problems. There is a lack of large, annotated datasets for training and evaluating Pavement Crack Information Extraction (PCIE) algorithms, which hinders progress in developing more accurate models.

3. Authors need to reduce the similarity rate to less than 20%, currently, it shows 27% which is not acceptable in academic journals.

4. What are the limitations of the proposed method? Does the method work well in lighting conditions (e.g., shadows, reflections), what about weather conditions (rain, snow, and fog) and surface conditions (dirt, oil stains)?

Comments on the Quality of English Language

ok

Author Response

Comments 1: [The abstract is too long and some included information part can be moved to introduction section.]
Response 1: We have appreciated your feedback and have streamlined the abstract to concisely summarize the core content of the paper.  We have moved more detailed information to the introduction to optimize the structure and enhance the reader's experience.  In the revision, we have outlined the limitations of 2D and 3D point cloud data in extracting pavement cracks and presented our proposed approach to address these issues.  The relevant details have been revised uniformly on lines 15 to 29 of the first page of the paper. 
Specific modifications are as follows："Please see the attachment."

Comments 2: [Authors need to provide detailed information about the training and testing process, and how to solve overfitting problems. There is a lack of large, annotated datasets for training and evaluating Pavement Crack Information Extraction (PCIE) algorithms, which hinders progress in developing more accurate models.]
Response 2: Thank you for pointing this out. In response to your questions about detailed information on the training and testing process, addressing overfitting issues, and the lack of a large annotated data set for training and evaluating pavement crack information extraction (PCIE) algorithms, we conducted an in-depth analysis and made improvements. We have added detailed descriptions of the sources of the datasets used, annotation methods, and tabular information. In addition, we plan to build and share custom datasets to support the development of more accurate models in the future, also added is a table for setting model parameters. The above has been updated on page 14 of the paper, 4.3 LNTU-RDD-Lidar Dataset, lines 595 to 600. Thank you for your guidance and support of our work.
Specific modifications are as follows："Please see the attachment."

Comments 3: [Authors need to reduce the similarity rate to less than 20%, currently, it shows 27% which is not acceptable in academic journals.]
Response 3: Thank you for your review and feedback on our paper. We note the issue you raised about similarity rates and fully understand the strict requirements for originality and similarity rates in academic journals. In order to solve this problem, we will carefully review and improve the content of the paper. We will reorganize and revise parts of the content to reduce similarity rates and ensure that citations and expressions meet academic standards. At the same time, we will use more advanced similarity detection tools to check to ensure that the similarity rate is reduced to less than 20%. Thank you very much for your suggestions and will complete these changes as soon as possible.

Comments 4: [What are the limitations of the proposed method? Does the method work well in lighting conditions (e.g., shadows, reflections), what about weather conditions (rain, snow, and fog) and surface conditions (dirt, oil stains)?]
Response 4: Regarding the limitations of the method you mentioned and its effects under different environmental conditions, we have carried out detailed analysis and supplement. The 3D laser point cloud pavement crack data has better robustness than 2D image under varying illumination conditions and low intensity contrast environment. It can effectively deal with various kinds of rust and oil covering on the road surface. In different weather conditions (rain, snow, fog), there may be certain limitations, especially in the case of water and snow cover can mask pavement cracks, making it difficult for laser scanning to capture the precise characteristics of cracks. To avoid these limitations, we will provide specific discussions and options on page 19, 5 Discussion, lines 694 to 701 of the paper in response to your comments and suggestions. Thank you for your guidance on our work and we look forward to your further feedback.
Specific modifications are as follows："Please see the attachment."

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The article introduces an innovative approach to detecting road surface cracks by utilizing advanced deep learning techniques, such as Point Attention Net (PAN) and the PC-Parallel module. This is significant because traditional methods often struggle with accuracy under challenging conditions. The article is a substantial contribution to the field of surface crack detection using 3D point clouds and deep learning. The introduced innovations, such as PAN and the Poly Loss function, show great potential in enhancing the precision and reliability of detection. This work can serve as a foundation for further research and improvements in this field, particularly in the context of automating the detection and analysis of road surfaces.

The conducted experiments demonstrated a significant improvement in metrics such as mIoU, Acc, F1, and Rec, confirming the effectiveness of the proposed approach compared to traditional methods.

The introduction of the Poly Loss function, which aids in better capturing the edges of cracks and details, is a well-thought-out solution that enhances the model's ability to recognize important features.

The complexity of the proposed model, including the attention mechanism and U-Net structure, may lead to higher computational costs. This can be problematic in practical applications where computational resources might be limited.

The method relies on 3D data, which can be challenging to obtain and requires complex processing. Furthermore, the paper does not address issues related to the disorder of point cloud data, which may affect the results.

The need for manual labeling of data to train the model is time-consuming and costly, which may limit the scalability and practical applicability of the method.

Although the method was tested on the LNTU-RDD-Lidar dataset, the paper does not analyze how the model performs under different weather conditions and on various types of surfaces. This may limit the generalizability and adaptability of the proposed approach.

The results of the experiments are promising, but the article does not provide sufficient information about potential obstacles in implementing the technology in real-world traffic scenarios, such as integration with existing infrastructure monitoring systems.

It is recommended to conduct further research on optimizing the computational performance of the model and to develop methods for automatic data labeling. Additionally, it would be beneficial to test the method on different types of road surfaces and under various weather conditions to assess its versatility and adaptability.

Detailed Notes:

Line 28 - Point Attention Net (PAN) - missing space; please check the entire article to eliminate punctuation errors.

Line 40: A period has accidentally been included before "Keywords."

Line 47: There should be a space before the citation “difficult[1].” Ensure that all citations throughout the paper comply with the journal's requirements.

Missing references to Figures and Tables, as well as incorrect references such as Table (1), etc.

Incorrect formatting in Equations 9 and 10.

Missing information on: Author Contributions, Funding, Institutional Review Board Statement, Informed Consent Statement, Data Availability Statement, Conflicts of Interest.

The paper should include a list of Abbreviations used in the work.

Author Response

Comments 1: [The complexity of the proposed model, including the attention mechanism and U-Net structure, may lead to higher computational costs. This can be problematic in practical applications where computational resources might be limited.]
Response 1: Thank you for following the discussion of model complexity and computational cost in our study. We highly value your comments that the attention mechanism and U-Net structure may lead to higher computing costs. To this end, we conducted a detailed evaluation of the number of parameters for the model and its comparison model. Our analysis shows that despite the increase in the number of parameters, our method still performs well in an environment with sufficient computational resources, and the number of parameters for each comparison model is increased in the 4.4 Comparative Experiment on page 15.In future work, we plan to further refine the proposed approach. By applying techniques such as optimal design and parameter sharing, we will strive to control the overall number of parameters to reduce the need for computing resources. Your guidance and support for our research is greatly appreciated and we look forward to your further feedback. 
Specific modifications are as follows："Please see the attachment."

Comments 2: [The method relies on 3D data, which can be challenging to obtain and requires complex processing.]
Response 2: Thank you for your review and feedback on our paper. We attach great importance to the difficulty of 3D data acquisition and the complexity of processing you mentioned, and hope to respond from the following aspects: Our research mainly contributes to the development of an effective use of 3D data to overcome the limitations of traditional 2D data under changing lighting conditions and low-contrast environments. Through in-depth study of three-dimensional data, our model is able to more accurately capture complex spatial structures and details, thus improving the accuracy of identification and analysis. Three-dimensional data provides richer spatial information than 2D data, which has significant advantages for recognition and classification tasks in complex environments. For example, in pavement crack detection, three-dimensional data can more clearly describe the depth and shape of the crack, thus improving the accuracy and reliability of the detection. In addition, we focus on building custom 3D datasets to support our research and provide valuable data resources for related fields. We have supplemented the paper with a detailed description of these aspects and introduced the collection and annotation details of the LNTU-RDD-Lidar Dataset on page 14, section 4.3 (lines 595 to 600), to more clearly demonstrate the value of 3D data and our research contribution. 
Specific modifications are as follows："Please see the attachment."

Comments 3: [The paper does not address issues related to the disorder of point cloud data, which may affect the results.]
Response 3: Thank you for your review of our paper and your valuable comments. We value your concerns about data disorder in point clouds and would like to share here how our approach addresses this challenge. Our research is based on the Pointnet++ architecture, which is specifically designed to deal with the disorder of point cloud data. Firstly, Pointnet ++ adopts hierarchical feature learning method to aggregate point cloud information layer by layer to extract local and global features of different scales while retaining data disorder. For downsampling, Pointnet ++ uses a neighborhood search strategy to select a representative set of points to effectively handle irregular and sparse point cloud data. Finally, through the multi-scale feature aggregation module, the model combines features of different scales to capture the spatial structure and geometric information of the point cloud, thus enhancing the robustness and accuracy of the model. We have further elaborated on these methods in the paper and introduced the details of the Point Attention Net processing point cloud data on page 8, section 3.1, Point Attention Net Model Overview. Thank you again for your guidance and support, and we look forward to your further feedback as we continue to improve this work.

Comments 4: [The need for manual labeling of data to train the model is time-consuming and costly, which may limit the scalability and practical applicability of the method.]
Response 4: Thank you for your review of our paper and your valuable comments. We fully understand your concerns about the time and cost constraints that can be imposed by manually labeling data. In response, we would like to elaborate on how our approach strives to address this issue. After we complete the initial training of the data set, the model can be inferred in practical applications only with unlabeled 3D data, so there is no need to manually label each time, which reduces the burden of labeling work to a certain extent. To further improve the scalability and annotation efficiency of the dataset, we plan to explore automated annotation like SAM and ChatGPT in future studies. These technologies will help us generate high-quality annotations more efficiently, thereby reducing the reliance on manual labeling and improving overall efficiency. We will add a scalability discussion of these improvements in Chapter 5 on page 18 of the paper, hoping that through these efforts, our approach will become more practical and scalable in the future. Thank you again for your guidance and support. We look forward to your further feedback so that we can continue to improve this work.
Specific modifications are as follows："Please see the attachment."

Comments 5: [Although the method was tested on the LNTU-RDD-Lidar dataset, the paper does not analyze how the model performs under different weather conditions and on various types of surfaces. This may limit the generalizability and adaptability of the proposed approach.]
Response 5: Thank you for your review of our paper and your valuable comments. As for the performance you mentioned in different weather conditions and various surface types, we have carried out detailed analysis and supplements. The 3D laser point cloud pavement crack data has better robustness than 2D image under varying illumination conditions and low intensity contrast environment. It can effectively deal with various kinds of rust and oil covering on the road surface. In different weather conditions (rain, snow, fog), there may be certain limitations, especially in the case of water and snow cover can mask pavement cracks, making it difficult for laser scanning to capture the precise characteristics of cracks. To avoid these limitations, We have provided specific discussions and options on page 19, 5 Discussion, lines 698-701 of the paper in response to your comments and suggestions. Thank you for your guidance on our work and we look forward to your further feedback.
Specific modifications are as follows："Please see the attachment."

Comments 6: [The results of the experiments are promising, but the article does not provide sufficient information about potential obstacles in implementing the technology in real-world traffic scenarios, such as integration with existing infrastructure monitoring systems.]
Response 6: Thank you for your review and valuable comments on our paper. We are pleased that the results of the experiment have been recognized by you, and we take seriously the potential obstacles you mentioned to implementing this technology in real-world traffic scenarios. In the current research, we are mainly concerned with the development and validation of the technology. However, we are aware that there are real challenges when integrating this technology with existing infrastructure monitoring systems. Including system integration compatibility to ensure that new technologies can be smoothly integrated with existing systems. The effectiveness of actual traffic scenarios, and the practicality of the technology is verified by real environment testing. Cost and maintenance, evaluate and optimize the implementation and maintenance costs of technology. We have add a detailed Discussion on the application of these real-world scenarios in the discussion section on page 19, 5 Discussion, lines 702 -710 of the paper to more fully illustrate the potential and implementation challenges of the technology. Thank you again for your guidance and suggestions, and we look forward to your further feedback.
Specific modifications are as follows："Please see the attachment."

Comments 7: [It is recommended to conduct further research on optimizing the computational performance of the model and to develop methods for automatic data labeling. Additionally, it would be beneficial to test the method on different types of road surfaces and under various weather conditions to assess its versatility and adaptability.]
Response 7: Thank you for your review of our paper and your valuable suggestions. We value your suggestions for optimizing the performance of model calculations, developing automated data labeling methods, and testing methods under different road surfaces and weather conditions. The following are the key projects in our future research. We plan to explore in depth how to improve the computational efficiency of the model. By adopting efficient algorithm design, we hope to reduce computational resource requirements and increase operation speed, thereby enhancing the feasibility of the model in practical applications. Second, we are aware of the limitations of manual data labeling, so we will work on developing automated data labeling tools. Using machine learning and computer vision technologies, we hope to achieve efficient and accurate data labeling, reduce the reliance on human labeling, and improve the efficiency and consistency of data labeling. The dataset is also expanded to cover different types of pavement and various weather conditions (e.g. rainy, snowy, foggy). This will help us more fully evaluate the versatility and adaptability of the model, verify how the model performs in real-world scenarios, and make necessary optimizations and adjustments. We believe that through these efforts, our approach will be better able to address challenges in practical applications. We look forward to further exploring these improvements in future studies. Thank you again for your guidance and support, and we look forward to your further feedback.

Comments 8: [Line 28 - Point Attention Net (PAN) - missing space; please check the entire article to eliminate punctuation errors.]
Response 8: Thank you for your review and careful feedback on our paper. We have carefully reviewed the entire article and made a comprehensive correction to all punctuation, including adding missing Spaces and correcting any other punctuation errors. 

Comments 9: [Line 40: A period has accidentally been included before "Keywords."]
Response 9: Thank you for your careful review of our paper and your valuable comments. We have carefully reviewed and fixed this error and conducted a thorough review of the entire article to ensure that no other similar issues exist. We appreciate your help in identifying these issues and we have be submitting an updated version and look forward to your further review and feedback. 

Comments 10: [Line 47: There should be a space before the citation “difficult[1].” Ensure that all citations throughout the paper comply with the journal's requirements.]
Response 10: Thank you for your careful review of our paper and valuable comments. We have examined the entire paper in detail and, according to the journal's requirements, made sure that all citations are properly formatted, including adding necessary Spaces and formatting adjustments.

Comments 11: [Missing references to Figures and Tables, as well as incorrect references such as Table (1), etc.]
Response 11: We have thoroughly reviewed all the graphic citations in the paper. We have corrected the incorrect citation format according to the formatting requirements of the journal, and we have added the correct citation in the corresponding text for the missing chart. Ensure that all chart references are compliant in the text. Thank you for helping us identify these issues in order to improve the quality and readability of the paper.

Comments 12: [Incorrect formatting in Equations 9 and 10.]
Response 12: We have carefully reviewed both formulas and have adjusted the layout of formulas 9 and 10 according to the format requirements of the journal to ensure that they appear correctly and clearly in the paper. At the same time, all formulas in the whole paper were thoroughly checked to ensure format consistency and compliance with the standards of the journal.

Comments 13: [Missing information on: Author Contributions, Funding, Institutional Review Board Statement, Informed Consent Statement, Data Availability Statement, Conflicts of Interest.]
Response 13: We note the lack of author contributions, funding, Institutional Review board statements, informed consent statements, data availability statements, and conflict of interest information that you have pointed out.  We have revised the paper based on your feedback and ensured that all required information is fully represented.  

Author Response File: Author Response.pdf