Edge HPC Architectures for AI-Based Video Surveillance Applications

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The manuscript benchmarks the performance of a variety of HPC architectures. The authors validate and interpret their results. 

 

However, it is also required to compare the results with similar works to highlight the importance of the proposed architecture.

Comments on the Quality of English Language

 Minor editing of English language required

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This article provides an interesting addition to the field through the readily available application (provided link to implementation). The manuscript largely falls within the thematic scope of the journal, but some considerations have to be addressed before publication:

• In abstract abbreviation, GPUs are used which is then explained multiple times e.g. lines 32, 269.

• Consider adding the YOLO keyword

• Fig. 1 - the flow of data is not clear - how do HPC and Edge Platforms relate, and where does AI take action?

• Repeated expansion of abbreviations (e.g. lines 22 and 28)

• Colloquial use of terminology - instead of “things” more suitable might be objects or entities

• Fig. 4 - It would provide more clarity to point terminology used in lines 163 (backbone), 165 and 174 (neck) in the picture

• Line 177 - hanging dot

• Fig. 4 - in the output section of the picture authors use variables that are not defined to describe dimensions of the tensor - bb and nc. Those variables are named after variables mentioned between lines 176 and 177. 

• Line 176 - “3-dimensional tensor” might correlate with the name on Fig. 4 - tensor of size [1,bb,nc+5], but between lines 176 and 177 authors provide 8-dimensional description. Usage of “+5” might indicate that this additional information is somehow hidden, but on the surface, it suggests that said dimension is expanded by a value of 5.

• Line 183 - again, usage of “+5” might be overloaded - does it add up to 85, or does it refer to additional information that is not kept as a dimension of the tensor

• Line 190 - usage of undefined class_score variable, this might also raise questions about the nature of the multiplication operator - is it normal multiplication of two variables or scalar multiplication?

• Line 198 - first declaration of abbreviation - IoU - should be in parenthesis, not after the comma since it was written so far this way

• Line 243 - The last 3 parameters of s are not explained

• In 3.2.2 Authors change the format of describing steps. So far, they used bullet points, now they use headings e.g. Detection, Estimation

• The article puts much emphasis on the theory and explanation of algorithms. It leaves 2 pages for the actual goal of the paper, which is hardware-related benchmarks, with one of those pages being used to explain benchmark setup. Readers can only assume that the software side was the same for each test since the authors did not mention that. There is also a need to mention how many tests were run, for how long they were tested and are those results from one test, all tests or average over all tests. The reviewer highly recommends the expansion of this section as well as adding a Discussion section which would allow to compare similar results and approaches. 

• The results section needs to discuss obtained results in detail and provide insight into why those results are the way they are. 

• The Conclusion section also does not provide what impact it has on Edge Computing systems that were introduced in the title. As far as the reviewer understands, the authors conducted experiments locally, which might yield different results in the case of Edge Computing systems. This issue should be addressed in more detail in the Discussion or Conclusion section

• Future directions or further study might be considered

• The code is available, but the authors probably don't make the dataset available. It might be helpful for further comparison and reproducibility of the results.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

Overall, this is a good submission. The problem is relevant and appropriate for the journal, the quality of writing and of illustrations is good, and the contribution is interesting enough. Reproducibility is good, both in terms of formal statements describing the novelty and in its summarization via algorithm boxes. There are a couple of issues which should be addressed however.

Firstly, the discussion of the results is overly brief and lacking in insight. There should be a lot more analysis and content here.

Another major objection that I have is that a lot of references are rather old. For example, in connection with AI surveillance some references are from 2012 which is ancient history in this field. More recent examples, such as Qian's 2022 work "Segmentation assisted U-shaped multi-scale transformer for crowd counting" (BMVC) or Wang's 2023 "Nwpu-crowd: A large-scale benchmark for crowd counting and localization" (PAMI) would be more appropriate. Other instances should be checked and updated as needed too.

 

Page 5: typo at the beginning of the first paragraph ". The" 

Comments on the Quality of English Language

Overall, this is a good submission. The problem is relevant and appropriate for the journal, the quality of writing and of illustrations is good, and the contribution is interesting enough. Reproducibility is good, both in terms of formal statements describing the novelty and in its summarization via algorithm boxes. There are a couple of issues which should be addressed however.

Firstly, the discussion of the results is overly brief and lacking in insight. There should be a lot more analysis and content here.

Another major objection that I have is that a lot of references are rather old. For example, in connection with AI surveillance some references are from 2012 which is ancient history in this field.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

Comments and Suggestions for Authors

My original review was positive and the authors have taken my feedback constructively addressing the issues I raised well. Hence I am happy to recommend acceptance.

Comments on the Quality of English Language

No problems here.