Face-Inception-Net for Recognition

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

For the face recognition task, this paper introduces a Face-Inception-Net composed of multiple FIN-Blocks that leverage a fusion of modern convolution and inception. Overall, the performance of this paper is promising, but there are some issues that need to be addressed.

1. In Table 2, the latest paper was published in 2019. Enhancing the comprehensiveness of the table by including more recent methods would be beneficial.

2. Transformer-based methods are missing from both Table 2 and Table 3.

3. Although the authors claim that the proposed method can address complex cases, these cases are not detailed in the experimental section.

4. This paper focuses on the face recognition task, with applications extending to related fields such as expression recognition [1], expression generation [2][3], and face reconstruction [4]. Exploring the relationships between these applications could provide insightful analysis in the introduction section.

[1] Facial Prior Guided Micro-Expression Generation

[2] Template inversion attack against face recognition systems using 3d face reconstruction

[3] Facescape: 3d facial dataset and benchmark for single-view 3d face reconstruction

Comments on the Quality of English Language

Fair

Author Response

Thank you for your review. I've made the necessary corrections as per your suggestions. I look forward to your prompt response!

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Overall, the flow of the paper is satisfactory, and the abstract is presented in a very organized manner.

The literature review that the authors have carried out is connected to the work that is being proposed.

The articles that are being reviewed are cited properly. 

Several CCN methods, as well as the benefits and drawbacks of various machine learning techniques, are discussed in the review section, which draws from articles that have been more recently published. 

In this study, they present a novel FIN-Block model that achieves a broad spatial receptive field and deep feature embedding extraction capabilities. This is accomplished by the utilization of a combination of modern convolution and inception. Hence novely is present in this paper.

Several learning factors, transfer learning, and hyper performance tuning parameters were highlighted by the author.

Excellent comparisons have been made between the suggested model and other existing models, which helps the author to pursue the work dynamically.

The way they concluded their research finding can be fine tunes some more better level of understanding.

Author Response

Thank you for your review. I've made the necessary corrections as per your suggestions. I look forward to your prompt response!

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

This manuscript proposes Face-Inception-Net (FIN) for face recognition, particularly for performance improvement in extreme scenarios and fairness situations. I summarize my concerns as follows:

My concerns are as follows: 1. The problem statements for extreme scenarios and fairness situations should further be elaborated since FIN is proposed to tackle these two cases. The definitions for both scenarios are not provided at least in the Abstract and Introduction, and why resolving these two scenarios are so important in practice? How FIN, e.g., with larger kernel sizes, can enhance performance in extreme scenarios and mitigate bias?

2. Although the experimental results are comprehensive, Section 3.3 requires a major revision. For example, the authors introduced ArcFace and CosFace, but Equation (7) which reflects the typical softmax is followed without explanation. Additionally, Equation (8) is perplexing, particularly regarding the definition of f(theta_yi). It does not align with ArcFace or CosFace, and the definition for the margin (m) is not provided. If m is defined as the minimum CosFace value across C subjects, the mathematical formulation should be clarified. In addition, proper justifications should also be provided for this definition.

3.  I recommend a thorough revision due to numerous unclear sentences that may require professional proofreading. Some examples are:

1. "These loss functions serve as supervision signals that promote the separability of features." >>> This statement is inaccurate; loss functions do not serve as supervision signals but the ground truth labels.

2.  

3. "Face recognition has achieved great results after years of development and technology accumulation. On the basis of these results we were able to conduct research. Neither the optimization of the model nor the application of the loss function is a simple task."

4. "However, since the development of hardware and the proposal of new technologies have never topped." >>> These sentences should be rewritten in a better way.
   
   4. I think the definitions of optimal value and average value in the experiment section should further be explained. The results summarized in Table 3 are not clear too, i.e., the table caption isn't consistent with the reported results.

5.  
6. 5. Given these issues, a major revision is advisable before acceptance consideration.

Comments on the Quality of English Language

Please refer to Item 3. 

Author Response

Thank you for your review. I've made the necessary corrections as per your suggestions. I look forward to your prompt response!

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Authors have solved all my concerns.

Author Response

Thank you for your comments！

Author Response File: Author Response.docx

Reviewer 3 Report

Comments and Suggestions for Authors

I recommend another round of mandatory revision. Additionally, it would be beneficial for this manuscript to undergo professional English proofreading.  One of the reasons is the excessive use of ornate language in the revision, which may not be suitable and could compromise the clarity of the manuscript.

For example, "In our experiments, after evaluating various loss functions[1,11], we selected ArcFace[1] as our definitive choice."
>> The use of "definitive choice", a phrase not found in the Oxford Dictionary. 

Another example, "ArcFace introduces a unified formula that blends the angular margin principles of both ArcFace and CosFace[11] with traditional softmax loss, providing a versatile framework for facial recognition.
>> Note that ArcFace (with an additive margin) and CosFace (with a penalized Cosine margin) are different! More specifically, ArcFace is NOT a unified formula for CosFace, to the best of my knowledge! 

Based on my previous comment:

Comments 1: The problem statements for extreme scenarios and fairness situations should further be elaborated since FIN is proposed to tackle these two cases. The definitions for both scenarios are not provided at least in the Abstract and Introduction, and why resolving these two scenarios are so important in practice? How FIN, e.g., with larger kernel sizes, can enhance performance in extreme scenarios and mitigate bias?

1.  For clarity, I suggest including examples of extreme scenarios and fairness issues associated with face recognition models, which could be in the Abstract or Introduction. 

2.  State the reasons why solving these scenarios is particularly important in practice.

3.  It would be beneficial for the authors to explicitly describe how and why the design of FIN contributes to performance gain. Specifically, elaborating why the adoption of large kernels, the stacking of large convolutional kernels, and the use of orthogonal kernels enhance feature discriminability in these scenarios would provide valuable insights to the readers.

4.  Referring to ArcFace and CosFace, please also describe how the representation vector of testing face images is extracted, alongside the feature dimension. 

Comments on the Quality of English Language

It would be beneficial for this manuscript to undergo professional English proofreading. 

Author Response

Thank you for your feedback. We have already polished the manuscript once, but the results still need improvement. We have carefully revised it again and welcome your critique and corrections.

Author Response File: Author Response.docx