You are currently viewing a new version of our website. To view the old version click .
MDPIMDPI
Search all articles
by
  • Seung-Beom Kang,
  • Seung-Gyu Kim and
  • Sang-Hyun Lee
  • et al.

Reviewer 1: Pyke Tin Reviewer 2: Anonymous Reviewer 3: Arbnor Pajaziti Reviewer 4: Artur Janowski

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Comments and Suggestion

1.    This paper presents a well-rounded and innovative approach to addressing the challenges in fishery resource management under the TAC system.

2.    The integration of RT-DETR and ARCore technology is a strong point of your study, showcasing how cutting-edge technologies can be applied to traditional industries like fisheries.

3.    The comparison between RT-DETR-x and YOLOv8x, with a clear presentation of mAP50 and recognition accuracy, effectively demonstrates the superiority of your proposed system.

4.    Provide more details on the experimental conditions, such as the diversity of the dataset, environmental factors (e.g., lighting, water clarity), and the specific hardware used for ARCore-based measurements. This will help readers better understand the robustness of your system.

5.    While your system shows high accuracy, it might be beneficial to discuss any limitations or challenges encountered, such as species with similar physical characteristics or the impact of external factors like poor lighting on measurement accuracy.

6.    Include a direct comparison of your system's performance with traditional manual survey methods to quantify the improvements in accuracy and efficiency. You may refer to the following references:

(i)            Thi Thi Zin, T. Morimoto, Naraid Suanyuk, T. Itami and Chutima Tantikitti, "Image technology-based detection of infected shrimp in adverse environments", Songklanakarin Journal of Science and Technology, 44 (1), pp.112-118, DOI： 10.14456/sjstpsu.2022.17, Jan. 2022.

(ii)          T. Morimoto, Thi Thi Zin, T. Itami, "A Study on Abnormal Behavior Detection of Infected Shrimp", Proc. on 2018 IEEE 7th Global Conference on Consumer Electronics, GCCE 2018, 2018, pp. 818–819, 8574860.

 

 

Comments on the Quality of English Language

The quality of English is fine. Minor checking will make it perfect.

Author Response

Thank you very much for your insightful comments, which have allowed us to improve the content of the paper. We have done our best to address your comments. 

We have attached our response as a file for your review.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

I have the following concerns.

1. RT-DETR uses ViT. In addition, since it is known that this model has better accuracy than YOLOv8, the conclusion given in the Abstract is known. What is your difference.

2.AR core is an augmented reality SDK for the platform. Android. Why? 

your difference

3.Euclidean distance(1) should be recorded for 3D coordinate.

4. As a rule, in the classic case for ML train, valid and test sets should be in the ratio 6:2:2. Therefore, comparisons with other models given in table 2 should be for the same conditions. This requires clarification.

5. As a drawback, the proposed model does not determine the weight and size of fish, which is very important in practical application.

Author Response

Thank you very much for your insightful comments, which have allowed us to improve the content of the paper. We have done our best to address your comments. 

We have attached our response as a file for your review.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

In this paper, the authors propose an automated system for fish species recognition and body length measurement, utilizing the RT-DETR (Real-Time Detection Transformer) model and ARCore technology to address these issues.

In addition, in paper is proposed the system that employs smartphone Time of Flight (ToF) functionality to measure object distance and automatically calculates the weight of Total Allowable Catch (TAC) managed fish species by measuring their body length and height.

The proposed model based on the obtained results reveal that the RT-DETR-x model outperformed the YOLOv8x 16 model by achieving an average mAP50 value 2.3% higher, with a mean recognition accuracy of 96.5% 17 across the 11 species.

Based on Table 2, Table 5, 6 and 7, the methodology and comparison of the obtained results is clear. Also description of the the proposed system for automating fish species recognition and body measurement, proves that the proposed algorithm indicate that faster and more accurate data collection is possible compared to traditional manual measurement methods.

The article is well written and composed, is a significant contribution to the efficient operation of the TAC system and sustainable management of fishery resources.

The references cited in the manuscript are appropriate and relevant to the research.

My suggestion regarding the improvement of the paper is as follows:

The text inside Figure 3 needs to be translated.

 

Comments for author File: Comments.pdf

Author Response

Thank you very much for your insightful comments, which have allowed us to improve the content of the paper. We have done our best to address your comments. 

We have attached our response as a file for your review.

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

While the paper presents an interesting attempt to automating fish species recognition and length measurement, there are several significant concerns regarding its real-world applicability, methodological limitations, and technical details:

  1. The system has been tested in artificial (laboratory) environments and may not account for real-world variables, such as water conditions, lighting, or fish movement in fish markets.
  2. Weight is calculated based on body length and height measurements, but the study doesn’t address potential inaccuracies due to variations in fish body shapes across different populations and their orientation within the bounding box (fish with elongated shapes rotated in the image could result in shorter length and greater height).
  3. Is it common practice to measure individual fish in real-world conditions? If so, the manual measurement process, along with electronic data entry, may not take much longer than using the system presented. However, in practice, I assume that measurements are taken from animals stored in nets or boxes, where numerous challenges arise—unresolved in this paper—such as rotation, occlusion, and overlapping.
  4. The paper mentions using basic data augmentation techniques, such as horizontal flips and rotations, but does not cover more advanced techniques like brightness adjustment, contrast changes, and occlusion handling.
  5. The ARCore system will likely underperform in poor lighting conditions or when there are moving shadows cast on the objects in the image.
  6. The system relies on a remote server connection, which seems problematic during fishery inspections carried out in the middle of the sea.
  7. The paper primarily compares the RT-DETR model with YOLOv8x only. Why only with YOLO not with eg.  Faster R-CNN?
  8. I believe a more effective solution would be to employ segmentation models rather than object detection models for image analysis, or to use a synergy of segmentation and object detection (classification).
  9. The training process was conducted for an arbitrary taken 300 epochs, but the paper does not discuss issues like overfitting finding etc.
  10. Validation metrics are lacking for example, confusion matrices would provide valuable insights into the model’s shortcomings and help identify potential areas for improvement.
  11. I would also appreciate a sequence diagram in UML format, rather than Figure 2 as it currently stands.
  12. The font size in Figure 1 should be increased. There is enough space in the diagram, and it’s difficult to read in printed version of the article.
  13. Table 1 is not an algorithm but an example of calculation. It would be better to present this in a more universal way, without assuming specific parameter values.
  14. In lines 181 and 182, it would be clearer to write "(x3,y3) is the 3D coordinate of p3," and the same for p4 "(x4,y4)." These changes should also be reflected in the basic formula (1). Furthermore, two coordinates alone do not define a 3D point.
  15. In the conclusion, it’s mentioned that the RT-DETR-x algorithm outperforms YOLOv8.x, but this does not prove that it is generally good. Moreover, the latest version of YOLO, YOLOv10, has improved detection capabilities for small objects in images.

 

Author Response

Thank you very much for your insightful comments, which have allowed us to improve the content of the paper. We have done our best to address your comments. 

We have attached our response as a file for your review.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

In general, I am satisfied with the answers to my concerns, except 1. The changes and additions made have improved the perception of the research conducted.

Reviewer 4 Report

Comments and Suggestions for Authors

Thank you to the authors for incorporating the revisions into their manuscript. Most of them demonstrate the authors' awareness of the issues that still need to be addressed, and they do not shy away from them. I am satisfied with the responses.