Enhanced Atrous Extractor and Self-Dynamic Gate Network for Superpixel Segmentation
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsSome of the Reviewer comments given below:
1.In the section 4.1 We follow 250 the same strategy as []. What strategy used to increase the no. of images?
2. Could you provide more details regarding the specific methods referred to in the text, such as ERS and ETPS, and how they perform in comparison to your proposed method?
3. In the context, when mentioning "low-level and mid-level properties," what specific properties are being referred to, and how does your method excel in segmenting these properties?
4. Can you provide more details about the specific components of your proposed method, such as the "Enhanced Atrous extractor" and "Self-dynamic gate," and their roles in the experiment on BSDS500?
5. In Figure 8, you mention the "baseline" and the "Enhanced Atrous extractor" as well as the "Self-dynamic gate." Could you clarify how the ASA (Average Segment Accuracy) values differ when these components are added or removed, and what this signifies for the effectiveness of each component?
6. When discussing the results of replacing the "Enhanced Atrous extractor" with other feature extraction methods in Figure 9(a), could you elaborate on the specific methods you compared, such as the "Vaillan transformer" and the "transformer of AINet," and the implications of their performance differences?
7. In Figure 9(b), where you explore the effect of different activation functions, could you explain the role of these activation functions in your "Self-dynamic gate" and why Softmax resulted in a significant performance decrease compared to Sigmoid, Tanh, or ReLU?
8. In Figure 11, where EAGNet is introduced to remote sensing images, what specific criteria were used to assess the accuracy of building segmentation in complex scenarios?
9. In Figure 10, you mention choosing images from the UCM dataset. Can you provide more information about this dataset, and how representative are these selected images of real-world remote sensing scenarios?
10.It's mentioned that the proposed EAGNet reduces the number of primitives. Could you clarify what "primitives" refer to in this context, and how EAGNet achieves this reduction?
11. Abstract can be more succinctly written. Abstract should be expanded sentences related to the results. The results of the study should be given as numerical percentages.
12. Can you summarize any novel contributions or insights presented in the research article that distinguish it from previous works on EAGNet?
13. Conclusions should be written in more detail adding numeric data.
Comments on the Quality of English Language
OK
Author Response
"Please see the attachment
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsTo improve the value of the algorithm you propose, we recommend testing it on image segmentation databases such as COCO 2017, Pascal Voc, ....
Please make the source code of the algorithm public
We recommend that you compare with image segmentation algorithms such as Unet, PANet, ResNet, PSPNet,...
Comments on the Quality of English LanguageNeed to correct spelling errors in the article
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsComments for author File: Comments.pdf
After minor corrections, the manuscript can be published.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 4 Report
Comments and Suggestions for AuthorsIn this work the authors propose Enhanced Atrous extractor, which introduces enhanced atrous convolution based on transformer architecture to extract multi-scale superpixel feature with contextual information. There are some comments to improve the manuscript:
1. it is necessary to add a small conclusion in the abstract section
2. There are several acronyms (SLIC, SNIC, GCN etc), can the authors add the meaning of this acronyms?
3. In line 85 there are two consecutive colon “Vision Transformer: Superpixel segmentation:”
4. In line 108 the sentence is repetitive “Gating mechanism: Gating mechanism
”
5. There are some missing spaces in line 112, 114. Can the authors check all the document?
6. After the state of the art, the contribution and differences with the related work should be remarked.
7. Sentence in lines 120-121 is unnecessary, instead of that, the authors can give a whole brief of the methodology section.
8. The figures are not cited in order (first figure 1 and then figure 4 for example) this causes confusion to the reader. Please reorder the figures.
9. Lines 135, 142, 146 there are unnecessary spaces ( review all the document)
10. In figure 2, it is necessary a better explanation in the caption (all the figures should be self-explained in the caption). It is the same for figure 3 to 12.
11. Check the format in lines 185-187
12. In section 4.1 a reference of the dataset is necessary. Also, there is a typing error inline 251. It section (4.1) should be re-written to clarify it.
13. All the libraries should have the version.
14. Give a detailed explanation of the metrics ASA, BR and CO. Also, in line 260 there are some missing spaces.
15. The font size in figure 7 is to small. The same in figure 10.
16. In figure 9 what is the meaning of the x-axis.
17. In figure 11, an explanation of what it can be seen is necessary
18. in subsection 4.7 there are several typing errors, missing spaces and the language can be improved.
19. A better conclusion it is necessary. It is important to resume the contribution, advantages and disadvantages, future work and applications.
20. The computational times can be interesting int eh comparison with other works, Can the authors add a comparison according the computational time?
Comments on the Quality of English LanguageThere are many typing errors, missing spaces and formatting errors, the language can be significantly improved. Please review the entire document to improve the quality of the document.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
Comments and Suggestions for AuthorsI have read your response to my questions. However, I find it unsatisfactory. COCO 2017, Pascal Voc 2007 and Pascal Voc 2007 are all databases used for object segmentation based on pixel-level segmentation.
CNNs such as Unet, PANet, ResNet, PSPNet are all CNNs used for object segmentation.
Therefore, your explanations are invalid.
Comments on the Quality of English LanguageThe article is not improved
Author Response
please see the attachment
Author Response File: Author Response.pdf
Reviewer 4 Report
Comments and Suggestions for AuthorsThank you for your attention to the comments provided. However, in order to improve the quality of the article and clarity for the readers it is necessary to address some points.
1. There are several acronyms (SLIC, SNIC, GCN etc), can the authors add the meaning of this acronyms? it is important for the readers to have all the information in the paper
2. After the methodology section, it is necessary to add more detail in the summary on lines 126-127 to provide readers with more information on the overall content of the methodology.
3. The computational times can be interesting in the comparison with other works, Can the authors add an explanation of why they were not calculated, as future work or a discussion around that?
Author Response
please see the attachment
Author Response File: Author Response.pdf
