AI Somatotype System Using 3D Body Images: Based on Deep-Learning and Transfer Learning

Round 1

Reviewer 1 Report

The research work about AI Somatotype System Using 3D Body Images: Based on Deep and Transfer-Learning is very interesting and valuable. I raised few concerns to improve the quality.

1.    Do you consider the topic original or relevant in the field? Does it address a specific gap in the field?

2.    In “Introduction” section, how does the contribution address a specific gap in the field? Please clearly describe the significant contribution in points.

3. It is suggested to discuss more related works in order to reveal the significance of the proposed research.   

4.   The article lacks vital discussion associated to the compared technologies. What are the other practical applications except smartphones?

5.       Figure 5 and 6 are unclear, please revise.

6.       In fig. 8, image augmentation block description should be revised. Also, explain figure 8 in the caption.

7.       How the authors compare to other transfer learning and deep learning models?

8.       In Table 5, there is undefined character in column 1.

9.       It is suggested to describe the Accuracy, Recall and Precision as equations from Table 6.

     The English language of the paper is understandable, however, there are areas where specificity could be improved to enhance the overall quality and impact of the paper.   

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

This study develops an AI body type system that is capable of using 3D body images collected through 3D body scanners to predict the three body types proposed by Heath-Carter's body type theory. My main concerns are given below.

1. Quantitative experimental results should be added to Abstract to enhance the attraction of the paper.

2. The motivation of the paper should be enhanced. The reviewer suggests the authors analyze the previous research and point out the motivation clearly.

3. The main steps of the method should be shown in Introduction. In addition, the contributions can be concluded in Introduction.

4. Currently, the authors use large number of samples for training, and only 10% data for test. Why do the authors use this partitioning guideline. Generally, we only use a small number of samples for training.

5. The font should be consistent. For example, Figure 6. the font of "Acc" and "Loss" is too large.

6. The architecture of the model is simple. The authors should show the architecture of the model in detail.

7. Table 5, the Korean font should be fixed.

8. Conclusion should be enhanced. Conclusions are not repeated descriptions of experimental results. Conclusion should point out the main findings of the research and conclude the shortcomings, thus provide the future research directions.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

The aim of this research is is to identify a body type: endomorph, mesomorph, ectomorph, by processing 3D scanner data. A 3D model of the subject is generated by the scanner. Three planar views are generated from this.  These are used as input to a CNN to classify the subject as one of the three body types. If the research is successful it could be extended to classify the thirteen detailed sub types. There is also a suggestion that a smartphone app could be created to do this task, although there are other factors that might make this impossible. 

The introduction gives a useful review of the topic an examination of why it’s useful, why it’s difficult and why their solution is so good. 

They present the methodology. The data set is a group of Korean males who gave informed consent. Reasons for excluding females were given. 217 males were sampled. To capture the data the subjects were dressed in skin-tight clothes. Eventually three planar projections were generated: of front to back, back to front and side to side.

They build the body type classifier using transfer learning, which is a sensible approach. 

There is a comparison for several NN architectures to determine the best for this task. Why were these specific architectures chosen? Having chosen one architecture it was trained on this task and the results evaluated. There is a reasonably sensible discussion of the results, although there is much emphasis on the training accuracy, whereas it’s the testing accuracy that is more important. The accuracy of the classifier is analysed using a confusion matrix. Ultimately, 27 samples have been used for testing, which is a small number. It would be instructive to estimate the errors in the accuracy, recall and precision.

Specific comments

Fig 1 and related text – why are these images excepted?

Line 99 - do the swim caps make any difference since the measurements aren’t close to the head?

Line 100 - why mention female’s attire since they were excluded from the study?

Line 118 – with a small data set, would you be better doing ten-fold cross validation?

Line 136 – so how was the ectomorph data augmented? Does flipping etc make sense given the symmetry of the views?

Line 155 – you say the simpler model gives the best results. Is this affected at all by the amount of data you have? i.e. would a more complex model perform better if you had more training data?

Line 179 – presentation of material is a bit odd. You mention the transfer learning and give results in section 2.4, then mention the architecture and hyperparameters in section 2.5. Are these parameters related to those of the previous section? Why? And again, fig 8 represents material that has already been explained.

Table 5 – some non-English text

Table 6 – is it possible to estimate the errors in the accuracy, recall and precision?

Line 287 – using a mobile phone is only good if the subject is willing to dress up in skin tight trousers!

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The authors have modified the paper according to the suggestions. I recommend to accept at this stage. 

Reviewer 2 Report

The authors have addressed all my concerns and I have no further comments.