Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Comments and Suggestions for Authors

Overall Assessment

The manuscript targets runtime sparsification for YOLOv4-tiny with the goal of lowering compute cost while maintaining detection quality. The study aligns with low-power deployment, yet the current version falls short of High level Q such as Q1 writting expectations due to limited benchmarking, insufficient theoretical grounding, and lack of hardware-level evidence. Improve English usage for precision and consistency.

Novelty and Positioning

Claims about IoU vs. F1 as a primary insight lack novelty; detection research already standardizes on IoU-based AP (COCO). Ground evaluation in COCO AP@[.50:.95] following Lin et al. 2014 (https://arxiv.org/abs/1405.0312) and anchor metric discussions with GIoU (https://openaccess.thecvf.com/content_CVPR_2019/html/Rezatofighi_Generalized_Intersection_Over_Union_A_Metric_and_a_Loss_for_CVPR_2019_paper.html) and DIoU/CIoU (https://ojs.aaai.org/index.php/AAAI/article/view/6999).

Methods, Matching, and Mathematics

Replace the ad-hoc Bounding Box Sorting Algorithm with principled bipartite assignment using the Hungarian/Kuhn–Munkres method; provide complexity, tie-breaking rules, and code pointers (Kuhn 1955: https://web.eecs.umich.edu/~pettie/matching/Kuhn-hungarian-assignment.pdf; Munkres 1957: https://www.math.ucdavis.edu/~saito/data/emd/munkres.pdf). Define all threshold functions explicitly, state variable domains, and add sensitivity analyses for sparsity factors with clear notation.

Datasets and Metrics

Expand beyond a small single-class set to multi-class COCO and Pascal VOC with AP@[.50:.95], AP50/75, and APS/M/L. Report mean ± std over multiple seeds, include confidence intervals, and perform significance testing. Provide per-category and per-object-size breakdowns to reveal failure modes under sparsification.

Empirical Validation and Hardware Evidence

Substantiate efficiency claims with measured end-to-end latency, throughput, and board-level power on an embedded platform. Report energy per frame under fixed power modes and identical preprocessing/NMS pipelines. Include retraining-free and structured pruning baselines for fair comparison, notably magnitude pruning (Han et al., NeurIPS 2015: https://arxiv.org/abs/1506.02626) and a channel-pruning reference; frame the contribution against a modern sparsity survey (Hoefler et al., JMLR 2021: https://jmlr.org/papers/v22/21-0366.html) and a directly relevant retraining-free baseline (Ashouri et al., Neurocomputing 2019: https://www.sciencedirect.com/science/article/abs/pii/S0925231219312019).

Figures, Tables, and Writing

Ensure each figure and table has a clear in-text reference, full captions, axis labels with units, and explicit column definitions. Add error bars and trial counts for curves. Standardize terminology, remove colloquialisms, and maintain consistent mathematical notation.

Literature Upgrade and Mandatory Citation

Expand the bibliography to 40–60 high-quality sources from 2022–2025 across CVPR/ICCV/NeurIPS/ICML and Q1/Q2 journals. It is recomanded to add a citation to Segal, J. Personalized Medicine 2023 (https://www.mdpi.com/2075-4426/13/5/874). Integrate the TSSCI sequence-to-image methodology with a Siamese similarity measure to assess temporal stability under sparsification, and place the citation in Methods for sequence aggregation and in Results for temporal-drift analysis. Suggested insertion: “We evaluate temporal stability by converting frame sequences into TSSCI composites and measuring Siamese similarity following Segal (2023), enabling sequence-level drift assessment under varying sparsity levels.”

Author Response

PFA Rebuttal Document 

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The author presented a new algorithm for the sparsity adjustments across network layers and weight distribution. Proposed approaches are novel and useful for the scientific community, hence the article can be accepted after significant modification.

Introduction section is very week and is not sufficient. 

The principle of the IQR method is missing; it should be explained with the data set dependency.

Hardware utilization is not explained in Section 5.

Author should conclude the work with the evidence in terms of the static data 

 

 

Author Response

PFA Rebuttal document

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

Dear Authors,

I read your article on the sparsification of the YOLOv4-Tiny model during inference with interest, and I would like to express my appreciation for the effort invested and the scientific value of the study. I am impressed by the original approach, particularly the introduction of the IQR-based sparsification strategy, as well as the development of the bounding box sorting algorithm. These ideas have clear potential for real-world application in resource-constrained environments and demonstrate a strong orientation toward the practical usability of the model. I also value the systematic comparative evaluation of different sparsification methods and the well-grounded justification for the choice of metrics. Although the article is well-organized and the results are presented with appropriate visualizations, I also encountered some difficulties related to missing definitions, unclear methodological specifications (e.g., the treatment of Non-Maximum Suppression), and incomplete information regarding the hardware used and the definition of fundamental concepts. In addition, I would appreciate more clarity regarding the role of the F1-score, which is initially dismissed, yet later used in the evaluation of the results.

Although the article, on the whole, leaves the impression of a well-executed piece of work, I nevertheless find a number of shortcomings. I have several comments and questions about them that I would like to discuss – they are described in the attached file.

Comments for author File: Comments.pdf

Author Response

PFA Rebuttal document

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Thank you for your submission and for providing a response to the initial review comments.

After careful consideration of your revised manuscript and response, I must respectfully assert that I cannot proceed with a full technical review at this stage due to the lack of necessary detail.

The primary technical obstacle is the inability to verify the claims of the changes made.

Your response to the previous comments is currently too general and does not sufficiently highlight the specific revisions within the text. To enable proper verification and a fair assessment of your work, I require the following:

1. Precise Identification of Changes: You must explicitly and methodically point out the exact locations in the manuscript where the changes were made.

2. Detailed Explanation: For every reported change, clearly state what the content was before the revision and what the content is now. This level of transparency is essential for the technical integrity of the review process.

Until these technical requirements are met and the revisions are clearly verifiable, I am unable to confirm the efficacy of your updates. I urge you to resubmit a revised package that rigorously adheres to these instructions.

Reviewer 3 Report

Comments and Suggestions for Authors

Dear Authors,

I would like to sincerely thank you for the work you have done on the article and for the thorough consideration of the recommendations provided. Your willingness to enhance the manuscript by including additional data, examples, and analyses has significantly enriched the quality of the presented research. Your work is an inspiration to readers and researchers. 

Best regards!