A Fast Instance Segmentation Technique for Log End Faces Based on Metric Learning

Round 1

Reviewer 1 Report

I have writtent the comments in the text.

Comments for author File: Comments.pdf

Author Response

We are grateful for the thoughtful and detailed feedback provided by the reviewer and have made revisions accordingly.

Q1: please move to conclusion

A1: In this version, we have made appropriate modifications to the Introduction. Thank you for your careful attention.

Q2: Please conclude how your study improves the previous studies? What is the novelty of your study? R-CNN and CNN have been used to measure end face of a log, according to the introduction

A2: In the Introduction of this article, various methods are compared. For traditional algorithms, there are issues such as the need for manual parameter tuning and the difficulty of simultaneous measurement for large-scale logs. For deep learning algorithms, taking the classic Mask-RCNN as an example, its main disadvantage is slow running speed due to the serial implementation of object detection and segmentation. Our proposed method can achieve parallel implementation of object detection and segmentation, which improves speed. In addition, the architecture utilizing metric learning effectively handles the relevant issues of overlapping regions between logs, thus improving the performance of the model. The specific modifications can be found at the end of the Introduction, where we have added a summary of the contributions to help readers better understand.

Q3: why four? I mean why not 3 or 5? If I understood correctly, you used a 4x4 convolution filter. If so, what index does the center pixel have? (w? h?)A tensor with odd dimensions, for example 5x5, should have been used

A3: I'm sorry that our previous explanation was not clear enough and caused misunderstandings. In this article, the detection model uses stacked 33 convolutional modules to downsample the original image to 1/4 of its size. Therefore, each pixel in the feature map outputted by the detection model represents a 44 region in the original image. This is equivalent to the detection model generating detection points uniformly on the original image with a stride of 4. In this context, the detection point refers to the feature map pixel. As for why a downsampling rate of 1/4 was chosen, it was based on a reference to CenterNet. In this version, we will elaborate on these related contents carefully.

Q4: Why different workers? If the worker was the same, wouldn't the photographer's error be the same in all the photos?

A4: In this paper, shooting with different workers can enhance the diversity of the data set, thereby verifying the applicability of the model. Because in a real environment, there may be different workers performing operations.

Q5: Were they all from the same tree species? Do you think that different tree species do not provide different results according to the level of different stages?

A5: Due to the limitation of the data set, these data are all from the eucalyptus logs, which is also one of the shortcomings of this paper. The trees of different ages and different tree species you mentioned will have an impact on the measurement results. We will actively consider this factor and do specific research on it in future work.

Q6: please move to methods/ I suggest that move this paragraph to method/ is it your results or the method that you have used to measure the diameter?

A6: In this version, we have moved the description you mentioned to METHOD.

Q7: did you capture at different height? for example, maybe the overlap error decreases if you capture at the height of load center,

A7: In this version, we conducted experiments at different shooting heights while maintaining a shooting distance of 3 meters. We were pleasantly surprised to find that the overall performance slightly improved when the shooting distance was closer to the load center.

Q8: what is your discussion when the capture distance was less than 3m?

A8: In this version, we have added a relevant discussion ”As can be seen from Table 5, when the shooting distance is less than 3m, the measure-ment error slightly increases. We suspect that this is caused by uneven focusing due to the close distance”.

Q9: The volume of a tree is a function of diameter and length. When you evaluated the diameter measurement, do you expect a result that is inconsistent with the diameter in the case of constant length log volume?

A9: The diameter we get will have an error with the real value. First of all, from the statistical results, the diameter measured by the algorithm is slightly smaller than the real value. Secondly, as the distance, height, and angle of the camera change, the error will change accordingly. This is also the main research content of our follow-up work.

Author Response File: Author Response.docx

Reviewer 2 Report

-The paper should be interesting ;;;

-it is a good idea to add more photos of measurements, sensors + arrows/labels what is what  (if any);;;

-What is the result of the analysis?;;

-figures should have high quality. ;;;;;

-text should be formatted;;;;

-please add photos of the application of the proposed research, 2-3 photos ;;; 

-what will society have from the paper?;;

-labels of figures should be bigger;;;; Fig. 2;;;

-Is there a possibility to use the proposed research for other topics, thermal imaging etc.;;;

"Thermographic fault diagnosis of electrical faults of commutator and induction motors"

-please compare the advantages/disadvantages of other approaches;;; 

-Conclusion: point out what have you done;;;;

-please add some sentences about future work;;;

-references should be from the web of science 2020-2023 (50% of all references, 30 references at least);;;

Author Response

We are grateful for the thoughtful and detailed feedback provided by the reviewer and have made revisions accordingly.

Q1: The paper should be interesting.

A1: Thank you very much for your interest.

Q2: it is a good idea to add more photos of measurements, sensors + arrows/labels what is what (if any).

A2: Thank you very much for your suggestion. We agree that adding more photos with clear labels and arrows to indicate the measurements and sensors would be helpful in improving the clarity of our manuscript. We have modified Figure 5 by adding additional images to provide further illustration.

Q3: What is the result of the analysis?

A3: Thank you for your careful review. We have made changes to the structure of the fourth section of the article. Specifically, we have moved parts that are unrelated to the experimental results, such as data introduction, to the Method section, making the Results section clearer and more objective.

Q4: figures should have high quality.

A4: Thank you for your careful review. We have redrawn the Figure 2 to make it clearer and more understandable.

Q5: text should be formatted.

A5: We apologize for our previous lack of adherence to writing conventions. In this version, we have made every effort to improve the writing of the article.

Q6: -please add photos of the application of the proposed research, 2-3 photos

A6: Thank you very much for your suggestion. We have modified Figure 5 by adding additional results.

Q7: what will society have from the paper?

A7: Accurate measurement of timber during transportation is crucial, and our work aims to achieve quick timber measurement to enhance the efficiency of the entire process. We have added specific details at the end of the article, highlighted in red font.

Q8: labels of figures should be bigger;;; Fig. 2;.

A8: We have redrawn the Figure 2 to make it clearer and more understandable.

Q9: ls there a possibility to use the proposed research for other topics, thermal imaging etc..! Thermoaraphic fault diagnosis of electrical faults of commutator and induction motors'

A9: In theory, the instance segmentation model proposed in this article can be adapted to any modality. However, when applying it to a specific modality, corresponding data should be used for training.

Q10: -please compare the advantages/disadvantages of other approaches

A10: In the Introduction of this article, various methods are compared. For traditional algorithms, there are issues such as the need for manual parameter tuning and the difficulty of simultaneous measurement for large-scale logs. For deep learning algorithms, taking the classic Mask-RCNN as an example, its main disadvantage is slow running speed due to the serial implementation of object detection and segmentation. Our proposed method can achieve parallel implementation of object detection and segmentation, which improves speed. In addition, the architecture utilizing metric learning effectively handles the relevant issues of overlapping regions between logs, thus improving the performance of the model.

Q11: Conclusion: point out what have you done

A11: In this version, we have made adjustments to the wording of the conclusion, with the first paragraph summarizing the innovations and contributions of this article.

Q12: -please add some sentences about future work

A12: In this version, the discussion on future work has been added to the third paragraph of the conclusion. Thank you for your careful review.

Q13: references should be from the web of science 2020-2023 (50% of all references, 30 referencesat least);.;

A13: The newly added references in this article are mainly from the fields of small object detection and dense scene object detection. The main purpose is to explore the difficult detection problems of small-diameter logs and densely stacked logs in future work.

 

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

Dear authors

Thanks for your review

I hope you publish more articles in the future.

Reviewer 2 Report

The paper is good enough to publish.