Deep Learning-Based Algorithm for Recognizing Tennis Balls
Round 1
Reviewer 1 Report (Previous Reviewer 1)
The article deals with the problem of improving the functions of intelligent tennis ball picker robot (ITBPR) using convolutional neural networks (CNN). 2 methods are described and it is shown that the method using CNN gives better results.
The new version of the paper is much better. But I still have some important recommendations.
Recommendations:
1. Figure 2 is not mentioned in the text, therefore it is not important from the point of the article and should be deleted.
2. Section 4:
The processing of the results in the table is nice, but the processing of the loss function results in Figures 11-17 is still inconsistent and confusing.
The images do not have a description of what is indicated on the horizontal and vertical axis. Different scaling is used, making comparisons difficult.
There is no legend, there are 2 curves in each figure and it is not clear what they represent (this is necessary so that the reader does not have to search in the text).
Captions are insufficient (everyone can see that there are curves in the pictures), what is important is what the curves express.
It is not necessary to include all 7 figures, Figures 11, 15 and 17 are important.
Author Response
Dear Editors and reviewers:
Thank you for your precious comments concerning our manuscript entitled “applsci-2062461”. Those comments are all valuable and very helpful for revising and improving our paper, as well as the important guiding significance to our researches. We have studied comments carefully and have made correction which we hope meet with approval. We ensure that all changes in the article are flagged using the "track" feature so that editors and reviewers can observe any changes. Also the references cited in the article are related to the content of the original manuscript.
We would love to thank you for allowing us to resubmit a revised copy of the manuscript and we highly appreciate your time and consideration.
Correspondence to:
Aiping Xiao
Beijing Forestry University
Beijing, 10083, China
E-mall: [email protected]
Author Response File: 
Author Response.docx
Reviewer 2 Report (Previous Reviewer 3)
The article has been revised to have a complete logic, clearly described the problems to be solved, and also obtained good experimental results. But I also want to make the following suggestions:
1. When comparing the performance of different algorithms, the comparison of detection speed (inference speed) should be added.
2. This article made statistics on the number of GPU cores during the experiment, but did not say to use GPU training when introducing the computer configured. I suggest that you clearly explain the parameters of the experimental equipment you use.
3. It is hoped that the author will improve the structure of the model in future work to achieve better results in the future.
Author Response
Dear Editors and reviewers:
Thank you for your precious comments concerning our manuscript entitled “applsci-2062461”. Those comments are all valuable and very helpful for revising and improving our paper, as well as the important guiding significance to our researches. We have studied comments carefully and have made correction which we hope meet with approval. We ensure that all changes in the article are flagged using the "track" feature so that editors and reviewers can observe any changes. Also the references cited in the article are related to the content of the original manuscript.
We would love to thank you for allowing us to resubmit a revised copy of the manuscript and we highly appreciate your time and consideration.
Correspondence to:
Aiping Xiao
Beijing Forestry University
Beijing, 10083, China
E-mall: [email protected]
Author Response File: Author Response.docx
Reviewer 3 Report (New Reviewer)
The authors propose an approach to picking up a tennis ball based on a CNN model. I appreciated it as a general idea,
But on paper, I have some remarks and questions:
Paper presentation side:
1) Reformulate an abstract
2) Figure 15 is not quoted at text level.
3) Same title for three sections: 3.1, 3.2, and 3.3
4) two pages to describe CNN, I find a few lines and references very sufficient; otherwise, we are in front of a majestic course.
5) Page 9 and line 285: "I find one, but I believe it is step three."
I recommend that you give yourself more time to better represent your paper.
Scientific side:
1) You used deep learning for a dataset of 1000 images, and I think it's not enough for such a model; try to increase your base by using data augmentation techniques.
2) According to Figure 9: I observe that the majority of your database has only one very clear ball to segment it; try to add images with two or more balls and also some obstacles with the same color as the tennis ball.
3) I can't find a comparative study with another deep learning model that can give a very significant result.
4) I want to talk to you about real time, but I can't find a study and a time comparison at the test level.
5) You require a more in-depth related work study.
6) Formula (2): You calculated the sum of 3 loss functions; did you use a normalization technique?
7) I notice that you keep using the words "optimization" and "model optimization." If possible, tell me how you optimized your model, or what level of time or parameter selection this optimization is at, and base your answer on which algorithm this optimization is based on.
Author Response
Dear Editors and reviewers:
Thank you for your precious comments concerning our manuscript entitled “applsci-2062461”. Those comments are all valuable and very helpful for revising and improving our paper, as well as the important guiding significance to our researches. We have studied comments carefully and have made correction which we hope meet with approval. We ensure that all changes in the article are flagged using the "track" feature so that editors and reviewers can observe any changes. Also the references cited in the article are related to the content of the original manuscript.
We would love to thank you for allowing us to resubmit a revised copy of the manuscript and we highly appreciate your time and consideration.
Correspondence to:
Aiping Xiao
Beijing Forestry University
Beijing, 10083, China
E-mall: [email protected]
Author Response File: Author Response.docx
Round 2
Reviewer 3 Report (New Reviewer)
thank you for your efforts and your response to all my comments. except try to give a final version well presented
This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.
Round 1
Reviewer 1 Report
The article deals with the problem of improving the functions of intelligent tennis ball picker robot (ITBPR) using convolutional neural networks (CNN). In the conclusion, the authors state that they give-up two classifier-based methods of tennis ball recognition: the color-and-contour-based method and the cascade classifier-based method. Since these are results that are supposed to show the benefits of the work, these methods should have a more detailed description in the text as subsections.
The solved problem is interesting, but the article needs to be thoroughly revised. I do not recommend publishing it in the presented state.
I appreciate that the authors work with their own dataset. It would be good to evaluate already existing algorithms solving this problem on this dataset.
Reasoning:
1. Section 2:
There are inaccuracies in the descriptions of basic neural network models. Figure 1 (the known image of a layered neural network) has no explained parameters at all, and the caption is insufficient.
L. 73-79: A neural network layer contains multiple neurons, not a single neuron.
Caption of Figure 2 (without explanation of blocks) is insufficient.
2. Section 3:
The section contains a broader discussion of the loss function. It is possible to consider that the loss function has the above 3 components. Since the loss function in CNN is the basis for the training (backpropagation) algorithm, the section should state how the weights in the layers of the network are trained.
In Figure 6, the text is illegible.
Divide the section into subsections according to the methods presented in the conclusion.
3. Section 4:
The processing of the results is confusing, even if there are a lot of images. Comparisons of the results with different parameter settings is problematic. Combining images and using colors could make it better.
The bbox is not marked in the figures.
Since the section evaluates losses, it would be good to describe in more detail how losses are calculated.
4. References:
There are no citations in the text to some of the references listed, which needs to be done.
Author Response
Dear reviewer:
Thank you for your precious comments and advice concerning our manuscript entitled “applsci-1993462”. Those comments are all valuable and very helpful for revising and improving our paper, as well as the important guiding significance to our researches. We have studied comments carefully and have made correction which we hope meet with approval. Revised portion are marked in yellow in the paper.
We would love to thank you for allowing us to resubmit a revised copy of the manuscript and we highly appreciate your time and consideration.
Sincerely yours,
Di Wu
Author Response File: Author Response.docx
Reviewer 2 Report
The paper simply presents the implementation of a deep learning model for object recognition, including collecting training data, pre-processing the data, and training the model. The work is very simple, all the knowledge presented and the methods are simple.
There is not enough novelty to be published in this journal. Some concerns are as follows
- Lack of novelty
- Lack of presentation of problems and solutions
- lack of related work
- Lack of discussion of experimental results and proposed solutions
- Only 6 references are cited in the paper (although 23 references are presented in the Reference list).
Author Response
Dear reviewer:
Thank you for your precious comments and advice concerning our manuscript entitled “applsci-1993462”. Those comments are all valuable and very helpful for revising and improving our paper, as well as the important guiding significance to our researches. We have studied comments carefully and have made correction which we hope meet with approval. Revised portion are marked in yellow in the paper.
We would love to thank you for allowing us to resubmit a revised copy of the manuscript and we highly appreciate your time and consideration.
Sincerely yours,
Di Wu
Author Response File: Author Response.docx
Reviewer 3 Report
1. Overview：
This paper mainly uses Mask-RCNN to recognize tennis balls, and uses the loss function theory of Mask-RCNN to obtain the best hyper-parameter of the training network.
A lot of experiments have been done in this paper, and good results have been obtained. But the article lacks certain innovation.
2.Main modification problems：
（1）The biggest problem in this paper is the lack of experiments on the detection accuracy and speed of the model, because only comparing the size of the loss function is too one-sided;
（2）The abstract of the article is too simple and needs to be improved and revised;
（3） The article does not use all the references listed by the author.
3. Minor modification problems：
（1）Format error in i1 in Figure1;
（2）Figure 2 is too monotonous. It is recommended to enrich the color;
(3) It is recommended to write the loss function in Formula (2) in detail.
Comments for author File: 
Comments.pdf
Author Response
Dear reviewer,
Thank you for your precious comments and advice concerning our manuscript entitled “applsci-1993462”. Those comments are all valuable and very helpful for revising and improving our paper, as well as the important guiding significance to our researches. We have studied comments carefully and have made correction which we hope meet with approval. Revised portion are marked in yellow in the paper.
We would love to thank you for allowing us to resubmit a revised copy of the manuscript and we highly appreciate your time and consideration.
Sincerely yours,
Di Wu
Author Response File: Author Response.docx
