Conditional Generative Adversarial Networks for Domain Transfer: A Survey
Round 1
Reviewer 1 Report
Referee report for "Conditional Generative Adversarial Networks for Domain Transfer: A Survey"
This paper reviews recent advances in study of neural networks.
This referee is not directly involved with the field, and thus can provide a view point of a reader studying the paper to get up to speed with latest research.
Overall, the paper is interesting and the current advances are well-presented,
except for occasional rough text, which is likely a result of too much editing - not surprising with a review paper.
For example, in the introduction the authors write:
"In recent years, the most important achievements are [1–5]."
It would make sense to state what those achievements are, and only then cite the literature.
Later on, page 7, there is a sentence beginning "In order to mining more cycle.." which does not read right.
I would recommend the authors review the paper once more before publishing, and it could also benefit from comments of native English speaker.
Also, there is unexpected blue text on page 9 - if this is a note from the authors to themselves for future editing, it needs to be removed or rewritten.
Tables on page 13 and 14 appear to be too wide.
None of these comments impact actual science discussed in the paper, but the paper will be better received if it had a more polished presentation.
Otherwise, fine paper, and I especially liked the well-executed diagrams.
Author Response
Comments and Suggestions for Authors:
Referee report for "Conditional Generative Adversarial Networks for Domain Transfer: A Survey". This paper reviews recent advances in study of neural networks. This referee is not directly involved with the field, and thus can provide a view point of a reader studying the paper to get up to speed with latest research.
Overall, the paper is interesting and the current advances are well-presented, except for occasional rough text, which is likely a result of too much editing - not surprising with a review paper.
For example, in the introduction the authors write:
- "In recent years, the most important achievements are [1–5]." It would make sense to state what those achievements are, and only then cite the literature.
Response: Thanks for your comments very much. We have explained them with red fonts. The revised part is : “domain transfer method has been applied into solving many practical problems, such as geoscientific inverse problems[1], maize residue segmentation[2], person re-identification[3], data augment in face recognition[4] and age estimation[5].”
- Later on, page 7, there is a sentence beginning "In order to mining more cycle.." which does not read right. I would recommend the authors review the paper once more before publishing, and it could also benefit from comments of native English speaker.
Response: Thanks for your comments very much. We have revised it with red fonts. The revised part is “The cycle transformation in CycleGAN and StarGAN is relatively simple, in order to make this process dig out more information.”
- Also, there is unexpected blue text on page 9 - if this is a note from the authors to themselves for future editing, it needs to be removed or rewritten.
Response: Thanks for your comments very much. This is our revised part before, and we have removed the color now.
- Tables on page 13 and 14 appear to be too wide.
Response: Thanks for your comments very much. We have checked and revised them all.
- None of these comments impact actual science discussed in the paper, but the paper will be better received if it had a more polished presentation.
Response: Thanks for your comments very much. We have checked the whole paper and polished the presentation.
Otherwise, fine paper, and I especially liked the well-executed diagrams.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The authors of “Conditional Generative Adversarial Networks for Domain Transfer: A Survey” perform a comprehensive survey of the use of conditional GANs in the field of domain adaptation.
The abstract is accurate.
Major pluses of the introduction section are that it is relevant and cites the most important literature.
However, there are important aspects of the paper that are missing for it to be considered a survey.
-
The study needs to disclose the bibliographic method, - since those are not given it is not possible to reason whether the tools used for the completion of the study are suitable and correct
-
The study needs to disclose the inclusion criteria for the referenced papers
-
Since a survey needs to be especially aware of the research landscape the study needs to relate itself to the surveys that are already published in the topic of GANs in transfer learning / domain adaptation -> define clearly in the introduction section how is it different and what makes it unique
-
A standard practice for survey papers is to briefly delineate the research in numbers - how many papers were inspected, how many were identified to fit the criteria, how many are included in the study
The inclusion of sections 2 and 3 is welcome and makes the paper more valuable to readers from outside the immediate field. For better ogranisation, both those sections and all their children should be subsectioned under an encompassing ‘Methods’ section.
It is not clear how the ‘semantic level’ was assigned in Table 4. While there is a description provided in lines 367-370, it is not supported by any reference, nor experimental results. It is also confusing as to why coloring in a line drawing would be at a different semantic level than converting sketch to a photo.
The manuscripts contains comments like “ Intuitively speaking, if CGAN is compared to the human brain, a person who has never been to the city but has seen countless rural sunrises and sunsets can hardly imagine what the urban evening looks like.”, which are not supported by any kind of research, and could simply be not true - in this particular case humans can easily imagine all sorts of sceneries which never existed - just look at concept art to any sci-fi intellectual property. Comments which are not supported by an adequate reference or by experimental results are a case of commenting beyond the scope of the article, and should be eliminated from a scientific paper.
The paper does not provide any closing discussion, and the conclusion section is merely an abbreviated abstract. The remaining problem and challenges section however is a big plus.
Minor corrections to the style of writing are necessary, as it is sometimes informal (for example, instead of “another big contribution”, “another major contribution” should be used)
Author Response
Response to Reviewer 2 Comments
The authors of “Conditional Generative Adversarial Networks for Domain Transfer: A Survey” perform a comprehensive survey of the use of conditional GANs in the field of domain adaptation. The abstract is accurate. Major pluses of the introduction section are that it is relevant and cites the most important literature.
However, there are important aspects of the paper that are missing for it to be considered a survey.
- The study needs to disclose the bibliographic method, - since those are not given it is not possible to reason whether the tools used for the completion of the study are suitable and correct
Response: Thanks for your comments very much. This paper is a research review to summarize and discuss the existing studies in the field of CGAN, to make people easily understand CGAN. Therefore, the bibliographic method is manual search and selection of appropriate literatures.
- The study needs to disclose the inclusion criteria for the referenced papers
Response: Thanks for your comments very much. The reference includes some published papers by SCI journals or conferences, widely recognized papers (or highly cited papers) in this field, and a few novel papers. And we have added the explanation in the introduction with red fonts.
- Since a survey needs to be especially aware of the research landscape the study needs to relate itself to the surveys that are already published in the topic of GANs in transfer learning / domain adaptation -> define clearly in the introduction section how is it different and what makes it unique
Response: Thanks for your comments very much. There are many surveys on GANs, but few of them concentrated on the topic of CGANs in transfer learning. Therefore, this paper aims to conduct a comprehensive survey on the research of CGANs. This paper mainly focuses on the development and application of CGAN. Finally, we also discuss the problems and challenges of CGANs, encouraging more researchers to study CGAN and apply it into more application fields. We have concluded these in the introduction with red fonts.
- A standard practice for survey papers is to briefly delineate the research in numbers - how many papers were inspected, how many were identified to fit the criteria, how many are included in the study
Response: Thanks for your comments very much. We have manually inspected almost 190 papers about CGANs, while some of them are not officially published or have similar ideas. Therefore, the final listed references are carefully selected by us.
- The inclusion of sections 2 and 3 is welcome and makes the paper more valuable to readers from outside the immediate field. For better ogranisation, both those sections and all their children should be subsectioned under an encompassing ‘Methods’ section.
Response: Thanks for your comments very much. We have combined them into one section 2: The development methods of CGAN.
- It is not clear how the ‘semantic level’ was assigned in Table 4. While there is a description provided in lines 367-370, it is not supported by any reference, nor experimental results. It is also confusing as to why coloring in a line drawing would be at a different semantic level than converting sketch to a photo.
Response: Thanks for your comments very much. In fact, line coloring is only a change in color pixels of an image, while converting a sketch into a photo involves a change in the overall style of the image, it is a change in the higher semantic level of the image. There is a big difference in the style of different paintings, such as realistic style, ink style, dark style, etc. Therefore, the style change involves more than just color change. The semantic levels of an image can also be divided as visual layer (low level), object layer (middle level) and concept layer (high level). The low-level semantics is the color, shape, texture and so on of an image; the middle-level includes the attributes and features of the image, which is the state of a partial object at a certain moment in the image; the concept layer is the high-level, which is the overall content shown by images to humans. We have revised the explanation with red fonts.
- The manuscripts contains comments like “ Intuitively speaking, if CGAN is compared to the human brain, a person who has never been to the city but has seen countless rural sunrises and sunsets can hardly imagine what the urban evening looks like.”, which are not supported by any kind of research, and could simply be not true - in this particular case humans can easily imagine all sorts of sceneries which never existed - just look at concept art to any sci-fi intellectual property. Comments which are not supported by an adequate reference or by experimental results are a case of commenting beyond the scope of the article, and should be eliminated from a scientific paper.
Response: Thanks for your comments very much. We have checked and revised it.
- The paper does not provide any closing discussion, and the conclusion section is merely an abbreviated abstract. The remaining problem and challenges section however is a big plus.
Response: Thanks for your comments very much. We have added some closing conclusions in Section 6 with red fonts, i.e., “This paper focuses on the research of CGAN in domain transfer, and introduces the development process of CGAN from the perspectives of: the principle of CGAN, the development of loss function, the model variants of CGAN, evaluation methods and application fields. These studies fully demonstrate that the CGAN model has great potential and research value.” “However, due to the short life of CGAN to date, it is still in the initial stage of development, and the relevant theories and applications are far from maturity yet. Meanwhile, we have also discussed and listed the future development directions of CGAN in Section 5. In summary, more investigations on CGAN are necessary for consolidating its development and we hope this paper will contribute to researchers interested in applying CGAN in different domains.”. The remaining problem and challenges section is our discussions and the possible future directions on CGAN. We have also modified some redundant parts, as follows:
Please see the attachment for more revised details.
- Minor corrections to the style of writing are necessary, as it is sometimes informal (for example, instead of “another big contribution”, “another major contribution” should be used)
Response: Thanks for your comments very much. We have revised it and checked the whole paper to make this article more formal.
Author Response File: Author Response.pdf
Reviewer 3 Report
Title: Conditional Generative Adversarial Networks for Domain Transfer: A Survey
Manuscript ID: applsci-1851974
Journal: Applied Sciences (ISSN 2076-3417)
A brief summary:
The authors give an introduction to the principle of Generative Adversarial Network (GAN) then focus on how to improve it to achieve better performance, and how to evaluate such performance across the variants. Afterwards, the main applications of CGAN in domain transfer are presented.
The paper focuses on some improvement aspects for CGAN and reviews most of classical models and architectures paper focuses on the application of CGAN in the field of the domain transfer.
Strengths:
- The article is written in an appropriate way.
- The paper is well-written, and presents comprehensive analysis results.
- The conclusions are justified and interesting for the readership of the journal.
Weaknesses.
- Motivation and novality should be mentioned in the introduction.
- Contribution of the work should be mentioned in the introduction.
Recommendations:
- The language is good. This does not preclude reviewing the paper again, linguistically.
- The authors are suggested to try to better show the contribution, motivation and novelty of the paper in the introduction section, as well as to give more elaborate discussion.
Decision:
- Major revision.
Author Response
A brief summary:
The authors give an introduction to the principle of Generative Adversarial Network (GAN) then focus on how to improve it to achieve better performance, and how to evaluate such performance across the variants. Afterwards, the main applications of CGAN in domain transfer are presented.
The paper focuses on some improvement aspects for CGAN and reviews most of classical models and architectures paper focuses on the application of CGAN in the field of the domain transfer.
Strengths:
- The article is written in an appropriate way.
- The paper is well-written, and presents comprehensive analysis results.
- The conclusions are justified and interesting for the readership of the journal.
Weaknesses.
- Motivation and novality should be mentioned in the introduction.
- Contribution of the work should be mentioned in the introduction.
Recommendations:
- The language is good. This does not preclude reviewing the paper again, linguistically.
Response: Thanks for your comments very much. We have checked the whole paper and polished the representation.
- The authors are suggested to try to better show the contribution, motivation and novelty of the paper in the introduction section, as well as to give more elaborate discussion.
Response: Thanks for your comments very much. The revised parts are marked with red colors in the introduction. As follows:
Contribution: “ Our main contributions can be summarized as follows:
- This paper first summarizes the researches on the CGAN in the domain transfer.
- This paper summarizes the CGAN from different aspects, such as the loss function, the model variants, application fields and so on, so that readers can easily understand the situation of current research.
- This paper discusses the remaining problems and challenges of CGAN and provides the possible future directions of development of CGAN.”
Motivation and Novelty: “With the assistance of CGAN, domain transfer can be easily applied into different application scenarios. Therefore, this article will review the problems of domain transfer based on CGAN. In fact, CGAN could also be used to another application fields, such as data augmentation, image inpainting and so on. While their principles are similar, so it is very necessary to make a review on CGAN.” and “At present, a large amount of surveys have summarized the models of GANs, but few of them concentrated on the topic of CGANs in domain transfer. Therefore, this paper will first summarize the researches on CGAN and further discuss the problems and challenges facing of it, aiming to encourage more researchers to develop CGANs.”
Author Response File: Author Response.pdf
Reviewer 4 Report
This manuscript presents the review of Conditional Generative Adversarial Networks (CGAN) for Domain Transfer. This manuscript is well presented and informative. However, several issues need to be clarified before it can be accepted for publication.
1. Why chose Conditional Generative Adversarial Networks (CGAN) as a method that was reviewed in this manuscript? Explain it in detail in the introduction
2. This manuscript only focuses on Domain Transfer. The authors need to explain this in detail.
3. Font color in lines 194-199 is blue. Is there any meaning to this?
4. Summary of several types of research has been presented in the table. But several of the only written in text. It would be better if all of the research are presented in tables or figures.
Author Response
This manuscript presents the review of Conditional Generative Adversarial Networks (CGAN) for Domain Transfer. This manuscript is well presented and informative. However, several issues need to be clarified before it can be accepted for publication.
- Why chose Conditional Generative Adversarial Networks (CGAN) as a method that was reviewed in this manuscript? Explain it in detail in the introduction.
Response: Thanks for your comments very much. We have explained the reasons with red fonts. The revised parts are “With the assistance of CGAN, domain transfer can be easily applied into different application scenarios. Therefore, this article will review the problems of domain transfer based on CGAN. In fact, CGAN could also be used to another application fields, such as data augmentation, image inpainting and so on. While their principles are similar, so it is very necessary to make a review on CGAN.” and “At present, a large amount of surveys have summarized the models of GANs, but few of them concentrated on the topic of CGANs in domain transfer. Therefore, this paper will first summarize the researches on CGAN and further discuss the problems and challenges facing of it, aiming to encourage more researchers to develop CGANs.”
- This manuscript only focuses on Domain Transfer. The authors need to explain this in detail.
Response: Thanks for your comments very much. We have explained the reasons with red fonts in the introduction. “Generally speaking, domain transfer is the process of moving data from one domain to another, by learning the similarities between these two domains and building a bridge between them, to apply the old knowledge into the new domain, so that the new knowledge can be learned faster and better. It can further address the image/scene gap when training and practically applying.”
- Font color in lines 194-199 is blue. Is there any meaning to this?
Response: Thanks for your comments very much. We are sorry for that. The blue part is the revised part before and no other meanings. We have removed the color.
- Summary of several types of research has been presented in the table. But several of the only written in text. It would be better if all of the research are presented in tables or figures.
Response: Thanks for your comments very much. We have added the references in table 3 and added a new table 5 to summarize the research of all application fields. As follows:
Please see the attachment for our revised table3 and table5.
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
The authors have addressed my concerns to a sufficient degree
Reviewer 3 Report
Required modifications have been done.
Reviewer 4 Report
Almost all issues have been resolved in the revised version. The paper can be accepted
