A Point Cloud Segmentation Method for Dim and Cluttered Underground Tunnel Scenes Based on the Segment Anything Model

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

 

This paper presents a method that segments a point cloud model of a tunnel scene using one of the existing segmentation methods, called SAM. After projecting 3D point clouds onto a 2D plane, manually annotated objects were utilized to segment a subset of a whole point cloud. Then subsequently, the method segments adjacent subsets based on the segmented subset point cloud. There are several comments that should be addressed to improve the quality of the manuscript.

 

1.       In the first paragraph of the introduction, it would be better to briefly mention why point cloud segmentation is significant in general or in specific tasks for establishing a solid starting point.

 

2.       In line 62, the second contribution is not clear. Is it something that SAM does not have but the proposed model has? Also, the third contribution looks similar to the first one.

 

3.       Additional explanations and descriptions are needed for determining tunnel slicing intervals due to its practicality. Similarly, the process of identifying an overlapping rate for adjacent slices is not clear. Is 10% from the trial-and-error approach? Is it a value that can be used in any other tunnel scenes?

 

4.       Is the proposed method scalable?  In the case of tunnels, it is expected that the scenes would be similar regardless of the length of the tunnels. However, due to the recursive nature of the proposed method, the scalability should be evaluated to ensure the performance presented via the experiment in this study.

 

5.       The most significant feature of the proposed method is a segmentation capability in dim and unstructured environments. However, the methodology does not specify how the method is able to effectively handle these conditions. What features the authors developed can handle such conditions? Which part of the method is handling these conditions?

 

6.       One highlight of the proposed method was generalizability, however, experiments and validation do not support this claim. To properly claim this feature, multiple rounds of experiments and validation would be required in different scenes.

 

7.       Manual annotations for the first frame can be further discussed when it comes to the performance comparison as it might be one of the reasons why the proposed method outperforms others.

 

 

 

Author Response

Dear reviewer，

We sincerely thank for your valuable feedback concerning our manuscript. These comments are constructive and helpful for revising and improving our manuscript. We have studied comments carefully and have made revisions which hope meet with approval. In the revised manuscript, all the changes are highlighted in red for easy inspection. 

For all response, please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This manuscript presents a concentre study on the tunnel model by a novel approach based on image processing and machine learning which is highly appreciated. In my opinion, it has the potential to be accepted for the publication because the authors’ concentration and direction are appropriated

A major revision has been suggested to this manuscript and it expected that all of them have to be well addressed in the revised version.

·       The language of the paper must be reviewed and conformed to the academic manuscripts, several dictation typos and grammar issues were found along the paper.

·       The literature review was talking about TLS technology, very short with no proper references cited, TLS is very useful method to scan the tunnel geometry, therefore, the Reviewer suggests the Authors to have a look on the following paper and include them in the TLS related works:

https://link.springer.com/article/10.1007/s00603-017-1166-6

https://link.springer.com/article/10.1007/s10064-015-0748-3

·       Additionally, there are other works focusing on the 3D Laser Scanning system to measure the geometrical changes and deformation in railway tunnels, coupled together, therefore, please look at the following papers, i.e.:

§  https://onlinelibrary.wiley.com/doi/full/10.1002/stc.2587

The former paper similarly deals with development of the 3D laser scanning on a tunnel including some different objects on the tunnel interiors, as you did in your work.  

 

·       A table of nomenclature is recommended in the revised version.

·       Section 2.2, Line 140, please cite a reference for Zhang.

·       It is highly recommended to use passive tense instead of active tense in the academic manuscripts, I mean, sentences start with “WE” should change to passive tense.

·       Line 200, section 3.1, when you are talking specific axis, it must be clear enough for the readers to understand the global coordinates, therefore, please clarify this matter by presenting an image.

·       Line 230, Figure 2, the Authors must clearly indicate why they have chosen 10% of the overlapping rate? Any study? Reference? Previous Works?

·        Section 3.2.1, the conversion matrix between pixel coordinates and world coordinates must be presented.  

·       The same comment for section 3.2.2.

·       Respect a unique decimal wherever presenting quantitative values.

·       Experiment: please include the camera and lens specifications used in the experiments.

·       What is the measurement accuracy of the results?

·       In the results processing, the whole geometry has been divided not different parts and then they have been merged together automatically, correct? There is a question raised in my mind, how did you handle this automatic process of merging? Have you made any type of indicator? Detector?  

·       When masking, what was the threshold?

·       Have you made any simplification on the obtained point clouds respecting the same geometry normal vectors?

·       Refereeing to table 2 and table 3, please calculate deviations and then include into this table, it will highlight the accuracy of your methodology.

·       Regarding the image processing, it would be great if you can present a flowchart describing your routines, algorithms developed within your calculation?

·       Please rewrite the conclusion section, it is poor comparing to what you have done in the paper. Please discuss the application, limitations and advantages of your work.

·       Regarding the computational costs, can you please discuss it?

Best Reagrds

 

Comments on the Quality of English Language

Please see my comments.

Author Response

Dear reviewer，

We sincerely thank for your valuable feedback concerning our manuscript. These comments are constructive and helpful for revising and improving our manuscript. We have studied comments carefully and have made revisions which hope meet with approval. In the revised manuscript, all the changes are highlighted in red for easy inspection. 

For all response, please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The authors proposed an underground tunnel 3D point cloud segmentation framework based on Segment Anything Model. The authors also provided experimental results comparing the proposed framework to previous works related classical region growing algorithms and PointNet++ deep learning algorithms.

Three contributions mentioned by the authors are as follows:

1. The proposed framework demonstrates a strong segmentation capability for tunnel point clouds in complex scenes.

2. The proposed framework avoids additional training and doesn’t depend on attributes like color and intensity

3. The proposed framework is tested via validation experiments involving 3D point cloud segmentation with complex settings.

 

The paper is well-organized and handles interesting issue. However, the proposed scheme depends its accuracy and performance on the existing SAM. The proposed framework changes a sliced 3D point cloud data to a 2D image and gathering the output from results of SAM. The authors explained the optimization or pre-processing method for applying to tunnel specific feature to fit the SAM.

The authors also considers bidirectional comparison of output results of the proposed framework.

The authors should compare or explain differences between the proposed framework and the following related works.

1. Tunnel Deformation Inspection via Global Spatial Axis Extraction from 3D Raw Point Cloud

https://www.mdpi.com/1424-8220/20/23/6815

 

2. Deep learning for large-scale point cloud segmentation in tunnels considering causal inference

https://doi.org/10.1016/j.autcon.2023.104915

 

3. Automated semantic segmentation of 3D point clouds of railway tunnel using deep learning

https://www.taylorfrancis.com/chapters/oa-edit/10.1201/9781003348030-343/automated-semantic-segmentation-3d-point-clouds-railway-tunnel-using-deep-learning-jeongjun-park-byung-kyu-kim-jun-lee-mintaek-yoo-il-wha-lee-young-moo-ryu

Comments on the Quality of English Language

Minor editing of English language required

Author Response

Dear reviewer，

We sincerely thank for your valuable feedback concerning our manuscript. These comments are constructive and helpful for revising and improving our manuscript. We have studied comments carefully and have made revisions which hope meet with approval. In the revised manuscript, all the changes are highlighted in red for easy inspection. 

For all response, please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

The reviewer agrees that obtaining accurate point cloud data and segmenting it is challenging due to the darkness and complexity of the mining tunnel. There's no doubt that the SAM algorithm, which recently emerged and is appealing in its zero-shot nature, is attractive. The authors evaluated the performance of the SAM algorithm in segmentation by applying it to an underground tunnel. In the reviewer's opinion, it seems that the SAM algorithm could be applied to utility tunnels or factories with numerous pipelines.

However, in the reviewer’s point of view, revisions seem necessary for the following aspects.

• Like most computer application-related papers these days, this manuscript also corresponds to a report on hands-on experience with an algorithm developed by other researchers. I would appreciate it if the conclusion mentions the contribution of this manuscript.
• The authors rotated the tunnel's Point Cloud Data (PCD) in the x-axis direction. It would be helpful if the paper discusses how to handle non-straight tunnels when applying the SAM algorithm.
• It would be helpful if there were an analysis of the quality of the Point Cloud Data itself to assess whether it was well built.

• 4.3. Comparative experiment
• The authors compared SAM with PointNet++ and Region-Growing algorithms, conducting an comparison analysis. It seems that data preprocessing was carried out for PointNet++ and Region-Growing, respectively. If the authors optimized their algorithm and the other two algorithms were not optimized in a similar manner, the results in Table 2 may lack meaningful comparison. It is crucial to ensure a fair evaluation across all algorithms for a valid comparison. For example, in the case of Region-Growing, a binary image was used. Did the authors use the same image as SAM, or did the authors convert it into a binary image with a different threshold value? Specific explanations are needed. If optimizing Region-Growing for binary images yields different results, it could be worth exploring.

• 5. Discussion and Conclusion

• Considering that SAM, a key component, is an algorithm developed by other researchers, it might be challenging to label this manuscript as 'novel.'
• While the introduction highlights SAM's strength as zero-shot, the conclusion stating an intention to improve it through deep learning seems inconsistent and warrants clarification.

Author Response

Dear reviewer，

We sincerely thank for your valuable feedback concerning our manuscript. These comments are constructive and helpful for revising and improving our manuscript. We have studied comments carefully and have made revisions which hope meet with approval. In the revised manuscript, all the changes are highlighted in red for easy inspection. 

For all response, please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

N/A

Reviewer 2 Report

Comments and Suggestions for Authors

Dear Authors

Very nice improvement, thanks!

Now Aceepted!

Best regards

The Reviewer

Reviewer 3 Report

Comments and Suggestions for Authors

The authors revised the paper as review comments. It is worth to accept.

Comments on the Quality of English Language

It is enough to understand.

Reviewer 4 Report

Comments and Suggestions for Authors

I've reviewed your revised manuscript and believe it is now suitable for publication in Remote Sensing.