Mixing Global and Local Features for Long-Tailed Expression Recognition

Round 1

Reviewer 1 Report

The authors proposed an expression recognition using a mixing of global and local features by using the long-tailed distribution data. The overall architecture of the proposed method is shown in Figure 3. The authors compared some datasets and methods to prove that their results were better than the proposed methods. Hope that the following comments and suggestions are provided to improve the article:

- The numbering of the reference needs to be ordered by number. It's hard to reference by the current sorting, I suggest rearranging the reference. For example, the first reference is [9] but not in [1].

- There are some missing references to the RAF-DB, AffectNet... Please check that all the methods are referenced properly.

- On page 4, line 184, the website should be written in reference.

- Figure 1 needs to be put on the same page as the first time reference in the paragraph. Moreover, Figure 2 needs to be referenced in the paragraph.

- There is a lot of space disappears, for example, on line 15 after "...88.92%", on line 69 after "...distribution[27]", and on lines 329 to 332, after the "...%" symbol. Please check all the typo carefully in the article. 

- In Table 2, the authors wrote "seven expression classes", but the items in the table have eight classes (maybe the "Contempt" need to be deleted?).

- In Tables 3, 4, 7, 8, and 9, compared with datasets and methods, maybe the reference can be typed in the Table.

- Some word needs a full name when first-time use in the article.

- Can you please provide the precision and result for the experimental result?

- Some English typos and styles should be double-checked.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

This paper introduces a method that combines local and global features for facial expression recognition. In addition, virtual samples created by fusing foreground and background images are augmented to training set in order to reduce the unbalance samples. Experiments were conducted in three datasets including RAF-DB, AffectNet-7, and AffectNet-8.

This paper needs proof-reading before submitting for review.

·      Some section titles (like 2.1, 2.2, …) uses the upper case but some other section titles (3.2, 3.3, …) uses the lower case.

·      Many fragment sentences like

o   Determine the weight of a channel …

o   Generate the final 2D …

o   Use the rectified linear …

·      Improper use of punctuations like

o   kernel size k. where …

·      Improper sentences like

o   The output weight co can be …

o   . . . as follows:       after alignment.

o   . . . 0.78%better

o   . . . ratio in subshttps://            (The authors should put the link to footnote or reference. This disturbs the flow of reading.)

o   . . . where M represents            (Here should not have indentation; it is not a new sentence.)

Citation is needed for Dlib algorithm.

The algorithm uses 68 key point positions, why does the summation is from 1 to 67?

The authors claim that one of the contributions is to combine local and global features for FER. The authors should also include the results for the 0 and 1, and 1 and 0 for local and global features. I bet the accuracies are similar to others. Thus, this is not really a merit and contribution. Besides, fusing features have widely used in the literature.  

The authors claim that to paste key pixels, overfitting is reduced. This needs justification.

The authors claim to achieve the most advanced experimental results. Some results are quite similar to other methods. Are they statistically significant? The author should show the amount of improvement that uses the cutmix data enhancement against the proposed algorithm without the cutmix data enhancement.

Data augmentation is not new. Through out the paper, I do not see any graph that discusses how long the tail is. All this paper addressed is the data unbalance problem; not really has anything dealing with the long-tail characteristic.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

The manuscript presents an interesting approach to dealing with the facial expression recognition problem. The work is interesting, but several points should be revised before acceptance. 

-        The paper’s organization should be added in the last part of the introduction.

-        The authors should revise the manuscript because it contains several overlapping between the texts.

-        When putting an equation, all its parameters should be defined. Please check all equations.

-        Comparison to some recently published papers under the same conditions (i.e., datasets + protocols) is necessary.

-        The conclusion section should be entirely revised. Make “occlusion” rather than “summary” and occlusion should not start with “in addition”.

-        Besides, the conclusion should contain the main results of this work.

-        The manuscript requires proofreading as it contains many mistakes.

-        References should be rewritten according to the MDPI format. 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

Thanks for the correction with my comments.

Reviewer 2 Report

Authors have addressed my concerns.

Reviewer 3 Report

The authors have taken all my previous seriously. For all these reasons, the manuscript can be accepted in its present form.