Pedestrian-Crossing Detection Enhanced by CyclicGAN-Based Loop Learning and Automatic Labeling

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper is entitled “Pedestrian Crossing Detection Enhanced by CyclicGAN-based Loop Learning and Automatic Labeling” In this case, the idea and results of the paper are interesting but the following comments can be utilized to improve this paper in future.

 

Abstract

• While the term "dual-loop generative learning" and "automatic image feature labeling" are mentioned, the abstract does not briefly clarify how these work or how they are linked to CyclicGAN. Briefly mention CyclicGAN and how it is integrated into the loop learning system.
• Phrases like “variations in contextual features substantially influence model performance” are too vague. Specify what "contextual features" refers to (e.g., weather, lighting, traffic density).
• Statements like “Experimental evaluations demonstrate…” can be more direct. Use active voice and name specific improvements, e.g., “The model achieved a 12% improvement in F1 score under low-light conditions compared to baseline methods.”
• The last sentence could more directly highlight how the results translate to real-world use. Strengthen by specifying “real-time driver warning systems” or “integration with in-vehicle assistance platforms.”
• The title references “CyclicGAN-based Loop Learning,” but the abstract does not explain what makes the learning “cyclic” or how it differs from standard GAN training. Add one sentence explaining the unique structure or benefit of the cyclic aspect in the learning loop.
•  

 

System Implementation

• Explain how the filtering of non-crossing frames was automated (if applicable) or done manually.
• LabelImg usage is standard—consider discussing any custom modifications or quality assurance measures in labeling.
• The weather imbalance analysis is strong, but quantify how many CycleGAN images were generated for each condition to balance the dataset (beyond percentages).
• Mention any failure cases or limitations of synthetic image generation (e.g., visual artifacts or mode collapse).
• Consider providing quantitative comparison of model performance with and without GAN-augmented images to highlight their utility.

 

Experimental Results

• While the section is structured logically (baseline performance, FID evaluation, performance improvement), there is redundancy in descriptions that could be tightened. Also, figures and tables are referenced, but critical interpretation of these visuals could be expanded.
• There is insufficient statistical analysis or comparative benchmarking against other augmentation techniques (e.g., random cropping, brightness adjustment). Without this, the reader is left to wonder whether CycleGAN truly adds significant value.
• The use of IoU threshold increments (0.5 to 0.8) is appropriate, but results are only discussed qualitatively. Including numerical results or a table of F1-scores for each condition at selected thresholds would be more informative.
• Figure 9 is referred to, but not critically analyzed beyond a brief trend summary.
• No standard deviation or confidence interval is mentioned, which limits the robustness of conclusions.
• FID values are reported, but their interpretation lacks depth.
• It is unclear why an FID threshold of 200 was chosen.
• The relationship between FID score and actual model performance is implied but not quantitatively linked.
• While results show trends, the experimental setup lacks clarity in how models were validated.
• Results are only reported for IoU = 0.5 — higher thresholds should also be evaluated to show robustness.

 

Conclusion

• The conclusion does not acknowledge any limitations of the current work. Add a brief statement acknowledging potential weaknesses such as reliance on synthetic data fidelity, or generalization limits of GANs across drastically unseen environments.
• The automated labeling module is mentioned briefly but not sufficiently emphasized. Discuss the novelty and efficiency gain of the automated labeling system more clearly—was there a quantifiable reduction in manual effort or time?
• The conclusion lacks mention of quantitative improvements (e.g., average F1-score improvements with synthetic images). Include a concise reference to the magnitude of performance gains observed due to the CycleGAN augmentation (e.g., “an average F1-score increase of X% in adverse weather conditions”).
• The conclusion currently reads more like a restatement of the results section without enough synthesis or critical insight. Strengthen the narrative by summarizing the key contributions in a more integrated way. For example, explicitly highlight how the integration of CycleGAN and automated labeling worked synergistically to address the dataset limitations and improve generalization.
• The future work is promising but described in general terms. Elaborate slightly more on the technical feasibility or roadmap for real-time video integration. What are the anticipated challenges (e.g., latency, model inference time, bandwidth)? Are there preliminary experiments or system architecture plans?

 

Final decision: This manuscript has interesting objectives; however, it needs MAJOR correction.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

I want to thank the authors for the submission. Pedestrians' safety is always a very interesting and challenging topic. The manuscript is well structured and written. However, I suggest authors to consider addressing the following: 

1. A few comments on the dramatic increase of the road traffic accidents, beside the numerical values are necessary to assist the reader to better understand the nature of the accidents and as well as the level of pedestrian's involvement. 
2. It would be useful if the authors provide information on why the 5 specific enviromental settings were only tested. 
3. Fig. 3 and 4 do not refer the digital background they use for a map as a source. 
4. The manuscript is interesting but the majority is spent in technical issues regarding the systems used. The authors' work is not framed around specific scientific questions, thus no research questions are found and of course answered. 
5. In general, I fail to understand whether the manuscript's contribution is methodological or just system intergation. 
6. Also, i havent found a clear "Limitations" paragraph or section, although there are quite enough of them (system trained and tested using data from a specific geographic area, no evaluation on cross-regional data or lublic benchamrsk, no actural error rate of false detections or annotation noise is presented, the framework is evaluated offline although future work is proposed to intergate real time video streaming, only five enviromental conditions are simulated, no cmoparison is made with other state of the art objec detectors justifying why YOLO was used, etc.)
7. The authors neither define the target audience or end users of their proposed system nor explain how they could use it. 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The paper introduces new knowledge and proposes a technical solution in the context of pedestrian crossing detection in difficult conditions (night, rain, or blinding).

I would like to see more detailed data on training parameters: number of epochs, batch size, loss functions used, and other settings.

The paper lacks a comparison with other pedestrian crossing detection methods, which would be valuable to show how the proposed solution compares to others.

The exact parameters of the YOLO model used in the study are missing. It would be good if the authors described in detail how they modified the original YOLO architecture, what parameters were set (e.g. number of layers, learning rate, batch size), and how they adapted the model to the specific conditions of working with pedestrian crossings.

The paper does not contain information whether the code implementing the algorithm and the datasets used are publicly available. It would be good if the authors presented information on the computational times needed to perform real-time detection, as well as hardware requirements.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

The paper regards pedestrian mobility. The authors propose a pedestrian crossing detection system that integrates CycleGAN-based data augmentation, and cloud-based automation to improve detection accuracy under varying environmental conditions.

The research issue is relevant and the paper is interesting.

However, the current version present some limits

The main limit regards the weakness in the consideration of mobility phenomenon studied in literature and modeled with Transport System Models. The methodology regards only the transport supply characteristics? Do you not consider the users and possible interaction winth pedestrian crossing? How it is possible to reduce road accidents without considering user's behaviour or pedestrian road flows? I suggest the authors to consider road traffic flow. For instance, recent literature propose the "estimation of a Fundamental Diagram with Heterogeneous Data Sources". Do you think that you can consider road traffic flows?

Please clarify if road traffic flows have a role in your analyses

Finally, please clarify if your methodology can contribute to "Sustainable Urban Mobility Plans: Objectives and Actions to Promote Cycling and Pedestrian Mobility". 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors have addressed almost all the comments.

Author Response

Thank you very much for taking the time to review our revised manuscript and for your constructive feedback. We sincerely appreciate your thoughtful comments and acknowledge the importance of addressing the core concerns you have raised regarding the manuscript’s innovation and scientific depth.

We are grateful for your guidance, which will help us significantly improve our manuscript in future submission.

Reviewer 2 Report

Comments and Suggestions for Authors

I would like to thank the authors for resubmitting their manuscript and for providing a detailed rebuttal letter. After carefully reviewing their responses, I have concluded that the revisions undertaken do not fully address the core concerns raised in my previous comments.

My feedback highlighted the need for a more in-depth and rigorous approach, recommending substantial revisions that could significantly enhance the manuscript's quality and contribution.

In its current form, the manuscript falls short of the journal's standards, particularly in terms of innovation and scientific depth. I therefore encourage the authors to further develop their work and consider resubmitting it at a later stage, once these aspects have been more thoroughly addressed.

Author Response

Thank you very much for taking the time to review our revised manuscript and for your constructive feedback. We sincerely appreciate your thoughtful comments and acknowledge the importance of addressing the core concerns you have raised regarding the manuscript’s innovation and scientific depth.

While we have made substantial efforts to incorporate the recommendations provided in your earlier review, we understand that there may still be areas that fall short of your expectations. We are committed to further enhancing the quality and contribution of our work. In our next revision, we will carefully re-evaluate our approach to ensure a more rigorous and in-depth analysis that meets the high standards of the journal.

We are grateful for your guidance, which will help us significantly improve our manuscript in future submissions.

Reviewer 3 Report

Comments and Suggestions for Authors

The authors have submitted a revised version of the manuscript, and it's clear that some improvements were made – the research objective is now more clearly defined, the explanation of the loop learning approach with CyclicGAN is more understandable, and the results section includes additional visualizations. These changes definitely improve the overall readability and presentation.

However, several important issues raised in the previous review have not been fully addressed, or were only partially covered. Here are the key points:

1. Basic settings like the number of epochs and batch size are now mentioned, but many crucial details are still missing – such as the loss functions used, optimizer type, learning rate values, and train/test data split. It would help to gather this information in a single table for clarity and reproducibility.

2. The paper still lacks any benchmark or comparative analysis with other pedestrian crossing detection methods. This is essential to understand how well the proposed method performs.

3. The manuscript states that YOLOv5 was used, but it's unclear which version or configuration was applied, whether any architectural changes were made, and if the model was trained from scratch or using pretrained weights. A few sentences clarifying this – or a simple schematic – would help readers understand the setup better.

4. There's still no indication whether the code or dataset will be made available. Also, no runtime performance data is provided – such as the time to process a frame, feasibility of real-time use, or required hardware specs. These aspects are important, especially for practical applications. Even rough estimates or a mention of availability plans would be useful.

1.  

Author Response

Please see attached file.

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

The new version partially adresseses my previous comments. I suggest to more underline the policy implications of research applications. The authors report some new senteces but they do not support these sentences with some references. I recalled some work in the previous review.  

Author Response

Please see attached file.

Author Response File: Author Response.pdf

Round 3

Reviewer 2 Report

Comments and Suggestions for Authors

Thank you to the authors for resubmitting a revised version of the manuscript and for addressing my overall comments.

While I believe there is still room for significant improvement, the manuscript in its current form is acceptable for publication.

 

Author Response

Thank you very much for your valuable feedback and for recommending the manuscript for publication. We sincerely appreciate your time and effort in reviewing our work and helping us improve the overall quality of the manuscript.

Reviewer 3 Report

Comments and Suggestions for Authors

The paper is solid and ready for publication, with just a few optional improvements to consider.
It might be helpful to make the system architecture (e.g., Figure 4) a bit more detailed — for example, by showing how pseudo-labels are created and filtered during the loop. That would make the process even easier to understand.
You could also consider adding a few recent papers (from 2022–2024) about pseudo-labeling or GAN-based object detection in road environments.

 

Author Response

Please see attached file.

Author Response File: Author Response.pdf