All reviews of published articles are made public. This includes manuscript files, peer review comments, author rebuttals and revised materials. Note: This was optional for articles submitted before 13 February 2023.
Peer reviewers are encouraged (but not required) to provide their names to the authors when submitting their peer review. If they agree to provide their name, then their personal profile page will reflect a public acknowledgment that they performed a review (even if the article is rejected). If the article is accepted, then reviewers who provided their name will be associated with the article itself.
All comments are addressed properly by the author.
[# PeerJ Staff Note - this decision was reviewed and approved by Xiangjie Kong, a PeerJ Section Editor covering this Section #]
Please consider reviewer comments on both this and the previous revision of the manuscript.
**PeerJ Staff Note:** Please ensure that all review, editorial, and staff comments are addressed in a response letter and that any edits or clarifications mentioned in the letter are also inserted into the revised manuscript where appropriate.
Check comment section
None of the comments have be addressed
No changes done
Thank you for your revision. However, upon careful review, it is evident that the majority of the comments and suggestions provided in the initial review have not been adequately addressed or incorporated into the manuscript. The responses appear to overlook or sidestep critical feedback rather than engaging with it constructively.
For the peer-review process to be meaningful and beneficial, it is essential that reviewer comments are taken seriously and reflected in the revised manuscript. I encourage a more thorough and sincere revision that demonstrates clear alignment between the feedback received and the changes made.
In their response, the authors have argued that many comments of both reviewers were already addressed in the manuscript. I suggest this indicates some issues in presentation, that readers are not identifying relevant material and I suggest revision to improve wording.
Experimental design
My earlier comment: Li 180 A suggestion is to keep one of the harvest populations (li 156, different locations) out of the training set to use as an independent test set.
Author Response: Thanks for your support and contribution. I don’t really understand your meaning. But the obtained raw images of fruit from different locations within the greenhouse were annotated and uploaded into the Roboflow tool. The images were randomly divided into train, valid, and test sets, expanded with diverse augmentation, and downloaded.
Further comment: The real-world issue with any model is 'how does it perform outside of its training set'. Here you have a structured data set involving different populations...but have mixed them and randomly subset to training and test. This is the best case performance scenario. In practice growing conditions can impact fruit and background canopy shape/colour etc, so the real world performance of the model is compromised. You had an opportunity to probe this with your data set by training/initial testing on n-1 populations, and then testing on the last harvest population, with it totally un-represented in the training set.
Validity of the findings
My earlier comment: My major criticism is that the differences in performance are modest, often <1% MAp, yet YOLO model development can be stochastic, there is a random element to it (li 252) Try training the same model 10 times and check the performance of these models on a test set …how much variation to you get….. if its 1% then the supposed improvement using FIS is really lost in noise.
Response: Thanks for your suggestion and contribution. The performance of a model does not only depend on accuracy but also includes speed, computation cost, and lightweight nature. The YOLOvFIS model was trained and tested as many times as possible. Additionally, its backbone, which replaces the backbone of YOLO variants, was also trained and tested many times before the reported results.
Comment: This response does not address the issue...the Abstracts claim is that the models perform better - but the difference (at <1% MAp) is not likely to be significant. I recommend remove the claim or proove significance. For example, if you train your model 20 times, because of random variations in sample order etc you will get slightly different models and thus different results on your test set. What is the SD of this variation, and how does it compare to your claimed increased performance between models?
**PeerJ Staff Note:** Please ensure that all review, editorial, and staff comments are addressed in a response letter and that any edits or clarifications mentioned in the letter are also inserted into the revised manuscript where appropriate.
**Language Note:** The review process has identified that the English language must be improved. PeerJ can provide language editing services - please contact us at [email protected] for pricing (be sure to provide your manuscript number and title). Alternatively, you should make your own arrangements to improve the language quality and provide details in your response letter. – PeerJ Staff
-
-
-
The author has proposed ‘YOLOvFIS: A YOLO Network for Fruit Instance Segmentation’. I have the following comments:
- Abstract: The number of images and classes is not mentioned in the abstract; the author needs to mention such details in the abstract.
- mAP and detection time are included, but metrics like precision, recall, or error rate are not presented individually.
- Add brief mention of key modules in the YOLOvFIS architecture (e.g., C4fR, SCDown).
- Introduction: However, the problem statement would benefit from a more concise and explicit summary. Currently, the challenges are described across multiple paragraphs, making it somewhat difficult for readers to quickly grasp the core problem being addressed. I recommend including a dedicated paragraph that clearly and succinctly states the main problem and its practical implications to strengthen the motivation for the proposed work.
- Contribution 4: Performance comparison is a result of the study, not a contribution by itself. It demonstrates the value of the contributions, but is not an original contribution.
- When it comes to recent studies author has not quoted recent studies done in the same area. Therefore, the author must enhance this section by quoting the latest studies.
- Some of the studies are as follows: the author should highlight their contribution.
- Fruit image classification model based on MobileNetV2 with deep transfer learning technique; Time-Sensitive Bruise Detection in Plums Using PlmNet with Transfer Learning; Optimizing Pear Leaf Disease Detection Through PL-DenseNet; Deep learning-based classification of alfalfa varieties: A comparative study using a custom leaf image dataset; Enhancing soybean classification with modified inception model: A transfer learning approach; Enhanced corn seed disease classification: leveraging MobileNetV2 with feature augmentation and transfer learning.
- Enhance the section by summarizing state-of-the-art comparisons.
- Consider adding a table comparing recent fruit segmentation studies with models used, datasets, and mAP.
- Dataset: It is good that the author has provided details regarding the dataset, but failed to show these details in tabular form.
- The author should mention details of all the classes in a table and then show a split of data for training, validation, and testing for each class before and after augmentation
- Mention the number of instances per class explicitly.
- Provide the reference for Roboflow and LabelMe.
- Model architecture: Architecture diagram provided, but consider including a schematic block diagram with flow of data from input → backbone → neck, → head.
- Results: Add confusion matrices for at least one key model comparison.
- Add cross-validation or justify if not done.
- Include an ablation study evaluating the impact of each architectural improvement.
- Add a paragraph discussing limitations and deployment challenges or opportunities.
- Add 2–3 lines specifying future directions, e.g., real-world field testing, generalization to other fruits.
-
This paper considers and adapts the YOLO series (nano and tiny versions) for the task of fruit instance segmentation. This is a valid topic for research. It also constructs an annotated image set, which will be a useful resource for comparative studies (if made available publicly)
Minor comments
Li 136: Why choose the C4fr module? Why choose nano models?
Wording needs to be clarified in multiple instances, for example follows:
Li 17, avoid colloquialisms like ‘beat’ here and elsewhere in the manuscript. Li 18 and elsewhere, check English, eg, ‘it lacked overall detection time’ needs rewording
Li 19, what version of YOLO do you refer to in YOLOvFIS, or is it an FIS adaptation of each version of YOLO that you compare?
Li 60, capital and italics for genus species
Li 247, clarification needed
Li 76-115 reads as a long but incomplete list of YOLO applications for fruit detection. Consider a Table.
Li 180 A suggestion is to keep one of the harvest populations (li 156, different locations) out of the training set to use as an independent test set.
My major criticism is that the differences in performance are modest, often <1% MAp, yet YOLO model development can be stochastic, there is a random element to it (li 252) Try training the same model 10 times and check the performance of these models on a test set …how much variation to you get….. if its 1% then the supposed improvement using FIS is really lost in noise.
29 figures are too much! Aim for <7
All text and materials provided via this peer-review history page are made available under a Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.