Accelerated Inference of Face Detection under Edge-Cloud Collaboration

Round 1

Reviewer 1 Report

The author written the manuscript very casually. The negligent starts from the title. Please change the

word “fetection” to “detection.”

1) The abstract is not readable.

2) There is no process time analysis, execution time, or time complexity analysis. So, I am not sure

how the author quantifying their study where the title indicates something else.

3) Figure 5 and 6 not justifiable. Need a little bit clarification.

4) Summary and future work should also talks about cloud security as Face detection model should

not be developed without considering the proper security concerns.

I hope the author will revise the manuscript to improve the quality of their work. The author should site

following papers in the literature review in order enrich the manuscript content.

1) Gupta, K. D., Ahsan, M., Andrei, S., & Alam, K. M. R. (2017). A robust approach of facial

orientation recognition from facial features. BRAIN. Broad Research in Artificial Intelligence and

Neuroscience, 8(3), 5-12.

2) Deng, H., Feng, Z., Qian, G., Lv, X., Li, H., & Li, G. (2021). MFCosface: a masked-face

recognition algorithm based on large margin cosine loss. Applied sciences, 11(16), 7310.

Best of luck with the work!!

Author Response

Dear reviewer:

Thank you very much for your valuable comments, we have revised it one by one, it is very helpful for us.

Author Response File: Author Response.docx

Reviewer 2 Report

This paper proposes a two-step acceleration strategy for the CenterNet object detection model. The proposed approach consists of a model pruning method that is used to prune the convolutional layer and the deconvolutional layer, and a neural network segmentation that makes full use of the computing resources on the edge and the cloud to further accelerate the inference of the neural network.

Overall, this is an useful study. Some details of the proposed approach need to be clarified.

1. Equation (1): Please clarify whether "l" represents the total number of layers or the layer index.

2. Equation (2): It would be better to use a single symbol to replace the current symbol "percent".

3. Please clarify how to solve the optimization problem defined in (4). In addition, the inference delay seems to be affected by the network transmission bandwidth (between the cloud and the edge). Is such a strategy adaptive or fixed for different networks?

4. Table 2: Please correct "easy, medium, and hard" to be "AP_easy, AP_medium, and AP_hard". Please also correct Table 3.

5. Table 4: Please add the unit of the delay values in this table.

6. Table 5: Please add the inference delay unit.

 

7. The experiments are conducted using the configuration in Table 1. Please discuss how the proposed approach can be generalized to be used for other configurations. 

Author Response

Dear reviewer:

Thank you very much for your valuable comments, we have revised it one by one. See the appendix for details. As regards sincere regards.

Author Response File: Author Response.pdf

Reviewer 3 Report

The authors introduces a two-step acceleration strategy for the CenterNet object detection model. This approach is shown to outperform existing techniques, and authors provide a comparison based on different metrics. A literature review is performed with many modern sources cited. The area is relevant, it is increasingly important now to be able to use face detection algorithms in modern technology.

The paper is well paced and structured, interesting to read, I have learned a lot while reading it. The method is clearly described, the results are well presented and necessary conclusions are made.

A couple of minor adjustments should be considered to improve the quality of the paper:

1)     Typo in the paper title – “face fetection”.

2)     It is necessary to provide more examples on where this method and similar technologies are used. Currently introduction section says that “face detection is widely used in today’s industry” but showing specific examples will provide more context to the paper.

3)     Tables 2 and 3 with the results should be improved by clearly showing, what the values mean. Currently, neither the table titles nor table columns/row tell what the values are. You can find in the text that these values represent the average precision values but that also should be clearly shown in the tables.

4)     Conclusion section should explicitly mention, where and how this new proposed method could be applied alongside the directions for future research.

Once this is amended, the paper will be suitable for publication in this journal.

Author Response

Dear reviewer:

Thank you very much for your valuable comments, we have revised it one by one. See the appendix for details. As regards sincere regards.

Round 2

Reviewer 2 Report

Thanks for your revision.

 

Point 3: You didn't clarify how to solve the problem statement defined in (4).

Point 7: You replied that "Although we only conduct experiments on limited hardware resources, this method is general and can be deployed on other hardware resources, such as Raspberry Pi or Tx2". Please justify this statement. Thanks.

Author Response

Dear reviewer：

   Sorry for replying you so late, your suggestion is great, I have revised the paper, see the appendix for details.

Author Response File: Author Response.pdf