Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The idea of deploying unmanned aerial vehicles to monitor road marking status is impressive and resolves multiple challenges that the conventional way faces. I would like to congratulate the authors on their great work. However, the manuscript has some issues to addressed before it is considered for publication. Below are my comments on those issues.

1. According to the guideline of Drones journal, the abstract has to have single paragraph. However, the authors presented their abstract in multiple paragraphs. Re-write the paragraph merging the paragraphs in to one paragraph.

2. The word "meanwhile" in line #68 is used unnecessarily.

3. Many abbreviation, such as the ones in line # 22 and line # 86 were used without definition. Define abbreviations on their first use.

4. image qualities such as that of Fig. 1 and 2 are very low. Try to enhance the quality (if possible)

5. In line #144, Table 2 with just a single line may not be necessary. The sensor specification can be stated in a single line in a paragraph. And, the same thing is true for Table 3.

6. In line # 310, the first statement is incomplete. 

7. Variables in Eq. 3 were not completely defined.

8. Equations need not be derived in results section. As shown in the manuscript Eq. 10 and on-wards have to be moved out of the results section.

9. In general, the authors are recommended to revise the manuscript in coherently.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Thank you for submitting your work that I found interesting to read. Here some comments for improvement.

1. You say "Data sharing not applicable. ". However, the text implies that you captured your own dataset. If so, this needs to be made available for other people to replicate and potentially improve your work.
2. Please improve the quaility of most of the figures. In many cases it was impossible to clearly read the legends.
3. Fig. 5: please correct to "Damage"
4. Did you think of using OBB annotations rather than the conventional rectangular bounding boxes (HBB). They should not be difficult to extract from your ground truth+some independent segmentation method like SAM
5. It is not clear if you compared your results on your dataset with those of other people. Also not clear if you tried your method on other people's data and compared results.
6. Please include a "Highlights" section as required by Drones.
7. I guess you do the processing offline, once you have obtained your image data. Please clarify, as the YOLO models you have used might be too heavy to use on-board.
8. Please report what values of hyperparameters you ended up using (again to help replication by others).
9. Please check your references thoroughly to ensure they are complete (if the paper is accepted, this is usually a major source of proof corrections).
10. Please show illustrative examples of failure modes

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The paper presents a quite well-developed methodology for road markings deterioration assessment, documenting very good performance at high TRL level. Limitations of the method (in particular in presence of sharp shadows) are discussed and should be the object of further developments, quite desirable for large-scale application, but not impairing the effectiveness of the methodology.

My main objection is about the fact that continuous or dashed lane lines were not considered. Of course, such lines are very important for traffic, therefore neglecting them is in my opinion a limitation of the methodology. I understand that long lines being extremely elongated, are not easily treated by bounding-box-based segmentation methods, but I think that the authors should justify their choice of not dealing with lines, and discuss the possibility of extending the framework to also include them.

The quality of figures in the pdf I reviewed is very bad, and sometimes it is very difficult to read them. Of course, I expect high-quality figures to be included in the final version.

Following, I list specific small issues that should be dealt with.

line 7 - Geoinformatic

line 44 - Specialized retroreflectivity vans: do you mean vans fitted with cameras? Maybe make it clearer

line 78 – anchor-free head: this expression is not clear to non-experts and in particular its implications are not evident, so if you mention that, you might add a few words to clarify. I would also say that “starting for YOLOv8” because then you use v9 and v10.

line 89 – (Remote Sensing): I suggest to remove this indication. Did you mean that it was published in Remore Sensing journal?

line 92 – spell out the acronym mIoU

line 100 – damaged.” remove fullstop

line 110 – “at the segment or instance level” this is not clear to all readers.

line 112 – “without explicit condition indices” do you mean that the deterioration is not quantified? Please clarify

line 130 – “in the city area” did you mean “in city areas”? You might as well say “in urban settings” or environments

figure 1 – maybe clarify that the map depicts South Korea (in fact maybe only a region of it? Gyeonggi-do?)

line 141 – what do you mean by spatial intricacy?

line 155 – the sentence seems incomplete. What is “with fewer than 60 pixels” referring to?

line 161 – several hundred pixels in area, or in both directions? 100 pixels is 1.26m: are indeed all markings several meters wide in both directions?

figure 2 – it is not clear what you mean by RM resolution and how it is measured. Is it the number of pixels comprised within the area of a typical RM?

line 195 – what do you mean by detector-centered patches?

figure 5 – Damage Ratio

line 221 - here you might discuss why such markings and not continuous or dashed lines

figure 6 - spell out fc

line 249 – again: clarify anchor-free heads and anchor shapes

line 257- “we train the detector on the split”: what do you mean? That you used the data as splitted into training, test and validation? Isn’t that obvious? “select the best ckeckpoint” what do you mean? Stopping criterion?

line 258 – when you first mention mAP50, mAP50-95 and F1 for the first time you should either explain what they mean or give a reference. “margin-padded crops”: clarify.

line 259 – clarify what you mean by centroid or mask centroid

section 3.4: you do not discuss the choice of hyperparameters of the UNet (number of levels, layers, down/up-sampling…)

line 305: what is a classical plug-in rule? In any case, implication of choosing a large or small h are not clear

line 306: what do you mean by “stabilizes mode structure”? Maybe that it finds the two modes reliably?

line 310: let p(x) [insert eq. (3) here] be the two-component mixture.

line 311 : define delta x

line 312 and following: This explanation of the EM algorithm cannot be understood without resorting to the cited paper. Eithe explain more clearly or just skip and let the reader go to the original paper. besides, many symbols in eqs. 4-7 are not defined, so that it is impossible to follow. And apparently, the procedure is iterative but the stopping criterion is not indicated. Also the largest posterior unnormalized score at line 318 cannot be understood without clarifying the whole method.

1. (9) clarify symbols, in particular the “1”

line 330 – since you do not spell out acronyms at line 329, use capital letters in line 330 to make the correspondence evident at line 330

section 4.1 – the discussion leading to the choice of YOLOv9e is very accurate and detailed (maybe even too much, in the economy of the paper), but I would comment that in any case the performance parameters are nearly equal for all models, so that other criteria might motivate the choice. For instance, the computation load and the speed of YOLOv10x appear significantly better

line 418 – spell out MAE

line 424 - Densenet

line 450 – I wouldn’t call these “failure cases” (maybe I would say so just of the last case of fig. 14). I would rather say that there are “minor localized failures”

line 460 – remove extra parentheses

figure 12 – clarify meaning of vertical lines

line 492 - missing space before “matches”

lines 492-493 – Notations are not used consistently, so it is not clear what is a method’s f and what fKDE because they are not explicitly defined above, or at least not with the same notation.

line 500 the Otsu threshold can be sensitive…

1. (14) – clarify notations. What is the 1? And then inside the sum is it a norm squared, or what?
2. (15) what is pi? and how are mu and sigma evaluated?

line 510 – This is not the continuation of the paragraph above, is it? You are adding an additional topic. Then, you should indent the line

line 519 – k-means

fig 13 (a)-(d) – use the same vertical scale for all figures

line 532 – first column

line 534 – remove “columns are… Error.”

figure 15 only shows very evident outliers that are easily removed by photogrammetric software as a routine pass, but does not show anything about sharper edges, etc. I think that this figure is useless and might be removed, or possibly substituted by a figure documenting “more uniform density and sharper edges”

line 554 – it is not clear what the reference model is and what the input pixels

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

Thank you for your revised manuscript where, in my opinion, you have addressed my questions satisfactorily.