Real-Time Ground-Level Building Damage Detection Based on Lightweight and Accurate YOLOv5 Using Terrestrial Images
Round 1
Reviewer 1 Report
- The introduction section lacks a narrative of the contributions of others in the field.
- The techniques used in the modification of the YOLOv5 structure are common in the field of deep learning. Can you add some of your innovations?
- In 5.2, the abbreviations of Ghost-CBAM-based YOLOv5 (G-YOLOv5) and Ghost-bottleneck-based YOLOv5 (G-YOLOv5) are the same, please modify.
- There are some grammatical errors in the text, please check and correct them.
- Figure 6 should have some adjustments.
- I have a question, can only use Ghost-bottleneck to replace the original trunk of YOLOv5 to achieve lightweight and improve the detection effect?
Author Response
Please see the attachment.
Author Response File: 
Author Response.docx
Reviewer 2 Report
Only a prototype would be missing, both code and data to be able to verify everything indicated.
My conclusion is that the publication of the article is accepted.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 3 Report
The authors have developed a new method for building damage detection using many images obtained in a field survey.
A large number of training data is created, verification is performed using multiple models, and the model with the highest accuracy and lightweight is adopted. From these things, I judged that the authors developed a method with high novelty and practicality.
However, I wrote comments about the points I noticed while reading the paper. I recommend you to fix it.
[LINE51]
Since it is not written what (a) (b) (c) in Figure 1 indicates, the symbol of each photograph should be added.
[LINE139]
The autohrs wrote that the field survey was conducted at two points, but it is possible that the damage will be repaired depending on the time, so I think that the date and period of the survey should be stated.
[LINE155]
The authors wrote that the size of the marked damage is 800x800 pixels, but looking at Figure 3, the size of the bounding boxes are not single. It is better to write the range of how big the marked damaged part is.
[LINE161]
The authors wrote that the dataset was divided into 70% for training and 30% for validation. How to divide it? Was it divided randomly?
[LINE168]
What is the size unit of figure4 (b)? Is it the square of the pixel?
[LINE266]
It is written "AS shown in Figure.6" but I think it is a "As shown in Figure.8".
[LINE274]
It is written "usingdifferent" but I think there is a missing as "using different".
[LINE354]
The authors not only classify the damage but also extract the damaged area, but I am wondering whether TP deals only with the accuracy of the classification or includes the area extraction.
For example, debri is distributed over a wide continuous area, and it seems difficult to detect the location.
From this point of view, the results shown in Table.2 seem to be fairly accurate.
The authors should clearly write the definition of TP in this paper.
[LINE368]
It is written "Table II" but I think it is a "Table 2".
[LINE407]
It is written "decrases" but I think there is a missing as "decreases".
[LINE439,460]
It is written "detetcion" but I think there is a missing as "detection".
[LINE443]
It is written "balanceing" but I think there is a missing as "balancing".
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Round 2
Reviewer 1 Report
The research has great engineering application value and the content is sufficient.
