Damage Detection and Segmentation in Disaster Environments Using Combined YOLO and Deeplab

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper proposes a  deep learning method for detecting the cracks and rubble in unclear building images. The findings seem to have some application values. However, I do not understand why the image is segmented after detection. As far as I know, many semantic methods can be selected to segment the image in one step. Some of the specific requirements  are given below.

(1) The conribtuions are not clear, please describe the fucus details in Introduction. 

(2) The proposed efficient trick just combines two traditional method. Why do not the authos use yolov9 and resnet? Please explain it. 

(3) It is also challenging to distinguish and understand the methods from the results. Also, the two steps use different GPUs. 

(4) More similar comparison algorithms (e.g. advanced deep learning) are needed to show the superiority of the proposed algorithm.

Comments on the Quality of English Language

Minor editing

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

I would recommend being more generous in technical explanations especially in terms of the use of networks by better specifying the type of training assumed and the number of datasets used. In some parts, which are too fragmented, one struggles to follow the objectives for which one is working.

Please clarify why: Segmentation is performed using Deeplabv2 only for cropped images.

Some typo such as : cracks. [16] proposed

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

Comments and Suggestions for Authors

 

Comments on Manuscript ID Remotesensing-3217697 entitled "Damage Detection and Segmentation in Disaster Environments Using Combined YOLO and Deeplab"

 1. In line 222 of the manuscript, DCIY is mentioned to "reduce unnecessary computations on the entire image," but there is a lack of quantitative data to support this claim. It is recommended to conduct experiments in terms of parameters, computational cost, and inference speed to quantitatively analyze the computational overhead of the three methods.

 2. The description in Section 4 "Experimental Setup" is not very clear. The introduction of the dataset in Section 4.1 is somewhat disorganized, lacking a central statement. In Section 4.2, the evaluation metrics and the description of the dataset collection environment are presented together, which is not logically coherent. It is suggested to move the description of the dataset collection environment to Section 4.1 to make the content organization more reasonable.

 3. The DCIY proposed in this study achieved better segmentation results compared to Deeplabv2 and InYD, but lacks comparisons with other advanced segmentation methods and performance under different conditions. It is recommended to add more quantitative experimental results, especially comparisons with the other segmentation algorithms, to further demonstrate the advantages of DCIY.

 4. It is suggested to optimize the design of the figures. For example, Figure 8 shows the comparison results of crack and pile detection, but the visualizations of InYD and DCIY are inconsistent. DCIY only shows the predicted segmentation results, which does not intuitively reflect the accuracy of the segmentation locations. It is recommended to keep DCIY's visualizations consistent with InYD to facilitate a more intuitive comparison of segmentation effects.

 5. In the discussion of limitations and future work, the manuscript mainly addresses issues related to the dataset but does not deeply explore the shortcomings of the model approach and improvement strategies. It is recommended to supplement the discussion on the potential weaknesses of the model and directions for future optimization.

 6. It is recommended to improve the overall language expression. For example, line 381 starts with "They are also affecting the evaluation metric," which uses a pronoun, making the expression less precise. It is suggested to condense the central statements of paragraphs to make the expression clearer, especially in Sections 3 and 4.

7. related works about Damage Segmentation should be introduced and discussed: https://doi.org/10.1080/10589759.2023.2234548

thanks

 

 

Comments on the Quality of English Language

none

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Most of the comments have been improved. But I still think yolov7 is not the best solution. If possilbe, the authors can explain and analyze it in some ways.

Comments on the Quality of English Language

Minor editing

Author Response

Please see the attachment.

 

Comments 1: Most of the comments have been improved. But I still think yolov7 is not the best solution. If possilbe, the authors can explain and analyze it in some ways.

Response 1: This result is the algorithm developed to operate on an actual exploration robot, and its performance has been verified through substantiation. In terms of the direction of the research, rather than applying a new version, we focused on improving performance by applying additional algorithms to the stabilized version and implementing it in real-time within the actual investigation robot operation system. Newer versions have been released rapidly, but applying them directly to the robot system was difficult, as it required more than just algorithmic implementation. However, even with the earlier version, the potential was sufficiently confirmed. Currently, with the latest version in use, we are exploring additional methods to improve performance, developing supplementary algorithms, and conducting verification.  We plan to write a new paper on this aspect at a later stage.

 

Minor editing: The overall English quality has been rechecked and highlighted in green.

Author Response File: Author Response.docx