Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The authors focus on the early detection of aggressive human behavior in real‑world video surveillance streams using a hybrid spatiotemporal deep‑learning model. They position their main scientific contribution in proposing an interpretable architecture that integrates motion analysis, temporal risk estimation, and spatial localization, thereby addressing methodological gaps left by existing video‑analysis approaches.

The dataset is described in detail: the authors select diverse video samples and then standardize them into a balanced and methodologically consistent corpus.

They conclude that, although the proposed architecture exhibits strong discriminatory performance under both partial and full video‑clip observation, these findings should be regarded primarily as evidence of its research potential for early detection rather than as sufficient justification for deploying it as a standalone tool in high‑stakes real‑world decision‑making scenarios.

It is also worth noting that their results confirm the model’s ability to localize not only motion but also semantically meaningful regions of aggressive behavior, which distinguishes it from standard classifiers that lack explicit spatial interpretability.

Moreover, the results highlight the approach’s practical usefulness for security‑oriented environments—such as urban monitoring, transportation security, and educational infrastructure—where rapid threat detection, robustness in complex scenes, and explainable automated decisions are essential.

However, the authors describe each stage of their work in great detail and justify their design choices very thoroughly, sometimes even excessively, in pursuit of what they call “a holistic approach to video data analysis.” In my view, some parts could be shortened—for example, lines 260–268 together with Figure 2 appear clearly redundant and do not add substantial information beyond what is already conveyed in lines 130–131.

As a direction for future work, when further developing the method, it may also be worth considering that Figure 5 shows intensive overfitting on the training set, which leads to worse validation performance compared with some of the other analyzed methods. This suggests that the parameters of the proposed model (as many as 1,203,791, as reported in Table 3) might benefit from simplification or better tuning. A similar tendency can be observed in Figure 5 for the CNN+BiLSTM model, where the large number of trainable parameters (173,699) also results in a substantial gap between the training and validation loss.

Although the results show that RAMT‑BinaryHeatNet achieves the highest final F1 score, there remains a considerable difference between the training and validation F1‑measures, which indicates room for further regularization and model refinement.

Author Response

We sincerely thank the Reviewer for the careful reading of our manuscript and for the insightful and constructive comments. We appreciate the positive evaluation of the proposed approach, particularly the recognition of its interpretability, methodological structure, and relevance for early detection in real-world video analysis scenarios. Below, we provide point-by-point responses to the comments.

Comment 1: Excessive detail and redundancy (lines 260–268 and Figure 2)
We thank the Reviewer for this valuable observation. We agree that certain parts of the manuscript were overly detailed. In the revised version, we have shortened the corresponding section and removed redundant explanations. 

Comment 2: Overfitting observed in Figure 5 and model complexity
We appreciate the Reviewer’s important remark regarding the observed training dynamics.  In the revised manuscript, we have explicitly addressed this point in the Discussion section. We have clarified that the observed train–validation gap reflects a residual overfitting tendency, likely associated with the relatively high number of trainable parameters in the proposed model (1,203,791 parameters) and in the CNN+BiLSTM baseline (173,699 parameters). We have also revised the interpretation of the training curves to avoid overstating model stability.

Comment 3: Gap between training and validation F1-scores
We thank the Reviewer for highlighting this important aspect. We fully agree that, despite achieving the highest final F1-score, the RAMT-BinaryHeatNet model exhibits a noticeable gap between training and validation F1-measures.

In the revised version of the manuscript, we have explicitly acknowledged this observation and incorporated it into the Discussion and Limitations sections. 

Once again, we thank the Reviewer for the constructive feedback. The suggested improvements have helped us significantly enhance the clarity, balance, and scientific rigor of the manuscript.

Reviewer 2 Report

Comments and Suggestions for Authors

This paper presents a well-designed and practically relevant approach to spatiotemporal classification of violent behavior in real-world scenarios. The methodology is clear, the experimental pipeline is thorough, and the comparison with multiple baseline models strengthens the validity of the results. The proposed hybrid model is particularly interesting and demonstrates strong performance across evaluation metrics, highlighting its potential for real-world applications such as surveillance and safety systems. Overall, this is a valuable and well-executed contribution to the field.

The only part that could be improved is the future work that could be described in more depth.

Author Response

We sincerely thank the reviewer for their positive and encouraging assessment of our work. We greatly appreciate the recognition of the methodological clarity, rigorous experimental design, and practical relevance of the proposed approach. We are also grateful for the recognition of the potential applicability of the RAMT-BinaryHeatNet model in real-world scenarios such as video surveillance and security systems.

Comment: Future research directions could have been described in more detail.
We thank the reviewer for this valuable suggestion. In the revised version of the manuscript, we have expanded the discussion of future research directions to present a more detailed and structured perspective. Specifically, we have addressed several key aspects in detail, including the development of more parameter-efficient model variants, the application of advanced regularization strategies, the extension of the framework to multi-class and multi-class behavioral scenarios, and evaluation on larger and more diverse datasets.

We also included practical considerations related to real-time deployment, such as latency optimization, implementation on edge devices, and integration into video surveillance decision support systems. These additions enhance the completeness of the manuscript and better reflect the broader research and application potential of the proposed approach.

We once again thank the reviewer for their constructive criticism, which helped us improve the quality and clarity of the manuscript.

Reviewer 3 Report

Comments and Suggestions for Authors

This paper proposes a spatiotemporal method for binary classification of violent/non-violent behaviors in real-world video scenes. The authors are suggested to prepare a revision to improve the submission. Some comments are listed as follows.

 

1. The article uses only a publicly available dataset containing 2000 videos from YouTube, resulting in limited scene diversity. It is recommended to supplement this dataset for validation.
2. It is recommended to design ablation experiments, gradually remove or replace key components, and analyze the performance of the proposed method.
3. The authors qualitatively presented several examples of location heatmaps, which can further provide quantitative evaluation results.
4. Please provide the standard deviation of the results from multiple repeated experiments to eliminate the influence of random interference.
5. Labe noise is an important issue in the field. The authors could discuss this. - Label recovery and trajectory designable network for transfer fault diagnosis of machines with incorrect annotation.

Author Response

We sincerely thank the reviewer for their constructive comments and valuable suggestions, which helped improve the clarity, rigor, and completeness of the manuscript. All comments were carefully considered and incorporated into the revised version. Detailed responses are provided below.

Comment 1:
This work uses only a publicly available dataset containing 2,000 YouTube videos, resulting in limited scene diversity. It is recommended that this dataset be expanded for validation.

Response:
Thank you for this important comment. The manuscript has been revised to clarify the scope and limitations of the experimental setup used. Additional text has been added to the Discussion section to clearly state that this study is based on a single publicly available benchmark dataset, which provides a controlled, balanced, and reproducible basis for comparing architectures.

Comment 2:
The authors provided several high-quality example location heatmaps, which can further complement the quantitative evaluation results.

Response:
We appreciate this valuable suggestion. The revised manuscript includes a quantitative assessment of localization performance. Specifically, the mean intersection over union (mIoU) metric was introduced to assess the quality of spatial localization.
Furthermore, the ablation analysis was expanded to include mIoU values, demonstrating how various architectural components affect localization accuracy. This provides a systematic and quantitative complement to the previously presented qualitative heatmap visualizations.

Comment 3:
Please provide the standard deviation of the results from multiple replicate experiments to exclude the influence of random noise.
Response:
Thank you for bringing this important aspect of experimental reliability to our attention. The manuscript has been revised to include measures of statistical variability. All key results are now presented as the mean ± standard deviation across three independent runs (n ​​= 3) with different random seeds (42, 73, and 101). This information has been explicitly added to the Results section, and standard deviation values ​​for validation metrics are now included in supplementary tables, providing a more robust and reproducible assessment of model performance.

Comment 4:
Label noise is an important issue in this area. The authors could discuss this further.
Response:
We thank the reviewer for raising this important issue. The revised manuscript now includes a separate discussion of potential label noise in publicly available video data.
It is explicitly acknowledged that real-world video corpora collected from open internet sources may contain annotation uncertainty due to ambiguous interactions, partial overlap, or hidden aggressive behavior. The manuscript now discusses how such factors can influence both model optimization and the interpretation of marginal predictions. Furthermore, several directions for future research were identified, including benchmark-based relabeling, consensus-based annotation refinement, and the application of noise-robust learning strategies.