DRE-Net: A Dynamic Radius-Encoding Neural Network with an Incremental Training Strategy for Interactive Segmentation of Remote Sensing Images

Round 1

Reviewer 1 Report

DRE-Net: A Dynamic Radii Encoding Neural Network with Incremental Training Strategy for Remote Sensing Images Interactive Segmentation

Please incorporate the below comments.

1.      Please state the database used in this study and the results achieved in the paper’s abstract.

2.      Paper organization section paragraph is missing.

3.      Figure 2 caption is too long.

4.      In the Related work section. The authors have cited the papers but did not discuss the results for each citation. What these researchers achieved in their work.

5.      What is the research gap?

6.       No section for a dataset. Although dataset details are present in the paper they should be in a dedicated section. Put a section after Related work and combine all the information about the dataset with some sample figures and tables.

7.      Proposed methodology section should have a flowchart.

8.      Sections 3.1 and 3.2 can have more sub-sections. Please number them.

9.      Please remove 4.1 from this section

10.   Section 4.2 should be a separate section

11.   Results section should purely discuss the results.

12.   More results are based on the comparison.

13.    No section for results.

14.   Make a new section for results and a separate section for comparison.

15.   In the present form, it isn’t easy to understand the results and contributions.

16.   Make a separate section for discussion

17.   The conclusion section is very week

18.   Make another section for Limitations and Future Scope.

 

 

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

This paper presents an interactive segmentation for remote sensing imagery. Through a few interactions, the segmentation accuracy such as IoU increased correspondingly. This paper is well written, and the experiments together with the comparison methods are clear. I think it’s a good paper, and I only have a few suggestions for it.

The used data are very high-resolution RGB images with three channel bands. Since the title mentioned that the method is for “Remote Sensing Images Interactive Segmentation”, I am expecting experiments with multi-spectral remote sensing images, which have multiple bands that may be mutually correlated. In addition, the segmentation is related to the scales of the image. Besides the very-high-resolution image, how does the proposed and comparison method perform when applied to a medium-resolution image, such as Sentinel-2? 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

The manuscript titled DRE-Net: A Dynamic Radii Encoding Neural Network with incremental Training Strategy for Remote Sensing Images Interactive Segmentation presents a method for improving error prone automatic image segmentation by incorporating manual intervention with only mouse clicks. I find this manuscript particularly interesting however, before I can recommend its acceptance some of the issues must be resolved.

 

I find the introduction and literature review to be satisfying. However, I find that numerous abbreviations tend to be confusing so I suggest that the authors only introduce the particular abbreviations when dealing with it, or at least add the abbreviation meaning once again when starting to deal with certain topic. Section 3 (proposed method) seems to be most confusing. It seems that the information is all around the text and to understand your proposal fully one must jump around the manuscript. So I suggest following:

1. Comment the Fig. 2 more thoroughly,

2. Add more details on 3. step in Algorithm 1 i.e. distanceTransforms. What is this particular step. What is it meant to do? How you define distance transform of click. Maybe even add some image.

3. You use the whole section to define thI  φ(min_dis, radii_list) function yet the function is never defined. Is this function same as DRE753? Is this function some generalisation of DRE753? If so your notation is lacking.

4. Maybe add how the interactive segmentation is used? Is this a two step process or multiple steps? User sees previous segmentation mask, then adds a bunch of clicks and the process is finished or the interaction continues until satisfying results? How much clicks can be added in step? Is it only one click or multiple clicks?

 

In the results section in Table 4, the last column is reserved for time, but I do not understand the format. Above is stated Time,s but it seems that format is H:M:S? Is this right? Can you also add what how much time is needed for Network to update results between clicks? This seems to be important if user continues to add click for refinement. If user needs to wait between updates for 15 minutes I hardly see your improvements as justifiable...

 

 

 

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 4 Report

This paper proposes an interactive segmentation model called DRE-Net and an incremental training strategy for remote sensing images. Reviewer’s comments are as follows.

1. Several words contain inconsistent uppercase letters, e.g., Interactive feature Maps in line 241 and train strategy in Fig. 8.

2. It is required to clarify the difference between the examples of (c) Disk Encoding and (d) DRE Encoding presented in Fig. 6.

3. Figures should be placed after being mentioned in the text.

4. The authors need to provide the explanation of ToR-Net in Fig. 8 and its reference.

5. The notation of γ in Eq. 4 is not defined in the text.

6. Table. 5 is not mentioned in the manuscript, and Table. 12 in line 444 does not exist in this manuscript.

7. The unit of the minimum distance in Fig. 5 is omitted.

8. Reference styles are not consistent with the instructions.

 

9. Conclusion should be improved to highlight the main contributions and weaknesses of the proposed methodology.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

The authors have addressed the comments

Author Response

Thank you for your valuable comments and helpful suggestions.

Reviewer 3 Report

Dear Authors,

thank you for revisions. I am satisfied with this revision and can propose acceptance of the paper.

Author Response

Thank you for your valuable comments and helpful suggestions.

Reviewer 4 Report

The authors conducted major revision according to the reviewer’s comments, and most of the concerns are resolved. However, there remains several issues in the current manuscript, and I suggest the following comments.

1. The revised manuscript contains meaningless characters “Y”.

2. It is required to modify “Table. X” into “Table X”.

3. The title of columns in Table 4 should be modified.

Author Response

Please see the attachment.

Author Response File: Author Response.docx