Deep Learning Optical Flow with Compound Loss for Dense Fluid Motion Estimation

Round 1

Reviewer 1 Report

Major comments 

1. The paper lacks some elements of good work, the reference are limited showing less work with less research  in deep.

2. There are no advanced machine learning models were developed. Those are existed models, however, probably not introduced for this engineering problem. Revise the title as per proper.

3. By looking at the abstract section, this section does not really present an abstract. You should follow those elements, i. problem statement or research motivation, ii. Main research aim, iii. Research results, general finding.

4. In the introduction section, you have explained the problem nicely, but you did not give credit for machine leaning to solve this problem.? I see no problem statement clearly presented as well.

5. Several Typos in this work please revise your entire work

Minor comments 

1. What is es L1 or L2 in the abstract 

2. This refrence is place wrongly .... e, and machinery manufacturing.[2]

3. .. eural network. [19]The... wrong also 

4.  No Reference in section 2.1

5.  Please put the theory before figure at 3.3. Analysis of RMSE Calculation Results

6. 

 

Author Response

Dear Editor and Reviewers：

We want to thank you for your valuable comments.

First of all, I am very sorry that this major revision took nearly 2 weeks. I hope to give some explanations: According to the valuable revision comments made by the reviewers, we have added some experiments, and experiments based on deep learning often require A lot of time to carry out, which led to our revision had to be postponed. Sorry for kept you waiting.

Next, I'll describe the modifications we made to the manuscript.

• First and most importantly, we have added Sections 3.2 and 3.3 to the article, respectively carried out parameter sensitivity analysis and validity demonstration of the composite loss function method at the core of this article, and made corresponding conclusions in the abstract, introduction, and conclusion
• We modified the title of the article to highlight the key points of the article according to the reviewers' comments.
• In Section 2.1, we added a table to illustrate the achievements of the FlowNet2 method in the field of computer vision.
• We further emphasize the innovative work done in this paper in the Abstract and Introduction.

These are the changes we made. For specific responses to your review comments, please see the attachment. Thank you again for your valuable comments on this article. For further information，please refer to the revised manuscript.Yours sincerely,Zhen Zhang

 

Author Response File: Author Response.docx

Reviewer 2 Report

Please see attached PDF for comments and suggestions for authors. 

Comments for author File: Comments.pdf

Author Response

Dear Editor and Reviewers：

We want to thank you for your valuable comments.

First of all, I am very sorry that this major revision took nearly 2 weeks. I hope to give some explanations: According to the valuable revision comments made by the reviewers, we have added some experiments, and experiments based on deep learning often require A lot of time to carry out, which led to our revision had to be postponed. Sorry for kept you waiting.

Next, I'll describe the modifications we made to the manuscript.

• First and most importantly, we have added Sections 3.2 and 3.3 to the article, respectively carried out parameter sensitivity analysis and validity demonstration of the composite loss function method at the core of this article, and made corresponding conclusions in the abstract, introduction, and conclusion
• We modified the title of the article to highlight the key points of the article according to the reviewers' comments.
• In Section 2.1, we added a table to illustrate the achievements of the FlowNet2 method in the field of computer vision.
• We further emphasize the innovative work done in this paper in the Abstract and Introduction.

These are the changes we made. For specific responses to your review comments, please see the attachment. Thank you again for your valuable comments on this article. For further information，please refer to the revised manuscript.Yours sincerely,Zhen Zhang

 

Author Response File: Author Response.docx

Reviewer 3 Report

This paper proposed a compound loss function of Euclidean distance error, angle error, and div-curl smooth loss to for Flownet2 in order to improve the accuracy of PIV. In general, the construction of the proposed loss function is physically defective. Considering its current status, this paper is not acceptable. The detailed comments and suggestions are listed as follows:

1.     .L(RMSE) (Eq. 1) is in m/s, L(AAE) (Eq. 2) is in degree, L(S) (Eq. 3) is in /s. The compound loss function L(Finals) was constructed by the sum of L(RMSE), L(AAE) and L(S). Since the unit is inconsistent for L(Final), the physical meaning of this compound loss function is not clear. The proposed compound loss function is physically not correct.

2.     The compound loss function can be constructed by the dimensionless version of Lrmse, Laae, and Ls. 

3.     The compound loss function can be constructed in the form of either sum or product. Either way should be rigorously derived with rationale.

4.     Thorough sensitivity analysis of the parameters should be carried out.   

Author Response

Dear Editor and Reviewers：

We want to thank you for your valuable comments.

First of all, I am very sorry that this major revision took nearly 2 weeks. I hope to give some explanations: According to the valuable revision comments made by the reviewers, we have added some experiments, and experiments based on deep learning often require A lot of time to carry out, which led to our revision had to be postponed. Sorry for kept you waiting.

Next, I'll describe the modifications we made to the manuscript.

• First and most importantly, we have added Sections 3.2 and 3.3 to the article, respectively carried out parameter sensitivity analysis and validity demonstration of the composite loss function method at the core of this article, and made corresponding conclusions in the abstract, introduction, and conclusion
• We modified the title of the article to highlight the key points of the article according to the reviewers' comments.
• In Section 2.1, we added a table to illustrate the achievements of the FlowNet2 method in the field of computer vision.
• We further emphasize the innovative work done in this paper in the Abstract and Introduction.

These are the changes we made. For specific responses to your review comments, please see the attachment. Thank you again for your valuable comments on this article. For further information，please refer to the revised manuscript.Yours sincerely,Zhen Zhang

 

Author Response File: Author Response.docx

Reviewer 4 Report

This manustricpt desribes the dense fluid motion estimation with improved Flownet2. It should be noted that, the flow simulations are now a modern tool and an integral part of the design process. Process simulation allows you to optimize existing structures, explore various new design solutions and the impact of changes on product characteristics. This article can provide an alternative approach for assessing and an optimization of the dense fluid estimation. 

The layout of the article is correct, both in terms of content and didactics. The article has been edited very carefully, and the style and linguistic correctness of the study are extremely correct. While reading it, I did not find any formal or content-related flaws and editorial errors. The figures included in the study fully reflect the essence of the analysed problems and allow for a complete understanding of the presented content.

However, I would like to point out that the publication lacks a list of symbols used in the text.

In my opinion, this paper can be published in Water, because it described issues and study are well aligned with the issue of this journal.

Author Response

Dear Editor and Reviewers：

We want to thank you for your valuable comments.

First of all, I am very sorry that this major revision took nearly 2 weeks. I hope to give some explanations: According to the valuable revision comments made by the reviewers, we have added some experiments, and experiments based on deep learning often require A lot of time to carry out, which led to our revision had to be postponed. Sorry for kept you waiting.

Next, I'll describe the modifications we made to the manuscript.

• First and most importantly, we have added Sections 3.2 and 3.3 to the article, respectively carried out parameter sensitivity analysis and validity demonstration of the composite loss function method at the core of this article, and made corresponding conclusions in the abstract, introduction, and conclusion
• We modified the title of the article to highlight the key points of the article according to the reviewers' comments.
• In Section 2.1, we added a table to illustrate the achievements of the FlowNet2 method in the field of computer vision.
• We further emphasize the innovative work done in this paper in the Abstract and Introduction.

These are the changes we made. Thank you again for your valuable comments on this article. For further information，please refer to the revised manuscript.Yours sincerely,Zhen Zhang

Round 2

Reviewer 1 Report

Accept in the present form 

Author Response

Thank you very much for your affirmation！

Reviewer 2 Report

The authors have well addressed the suggestions and comments. The paper is now ready for publication. Congrats and good job!

Author Response

Thank you very much for your affirmation！

Reviewer 3 Report

Considering its current stautus, I have to decline this paper.

As pointed out by the paper recommended by the author, "Kendall A, Gal Y, Cipolla R. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2018: 7482-7491", the naive weighted sum method "Loss = ⅀wi * Li" has a lot of issues, one of the issues is the difficulty of optimal weights, but this study still used this simple weighted sum method. 

 

Author Response

Thank you so much for your comment!Compound loss function may be a simple method, but it is found in the research of this paper that it can still help the model achieve better results.At present, some studies have shown that the use of compound loss functions can help optical flow estimation of fluid motion.

Yu C D, Fan Y W, Bi X J, et al. Deep particle image velocimetry supervised learning under light conditions[J]. Flow Measurement and Instrumentation, 2021, 80: 102000.

But this method itself is indeed as you said, there are problems such as difficulty in parameter adjustment and difficulty in model training convergence.

We have now indicated the limitations of our method in the concluding section and are ready to attempt to address this issue in the next phase of our research.