Anthropometric Ratios for Lower-Body Detection Based on Deep Learning and Traditional Methods

Reviewer 1

Comment:

Response to comment:

The paper is reasonably structured, and the methodology is appropriate.

Thank you so much for your valuable comments.

The evaluation is based on measures of accuracy, specificity and sensitivity. The choice of CNN architecture is not justified; one potential way to justify the choice is to use an ablation study, such as removal of a layer or changing its parameters and see how it affects the overall performance.

We are grateful for the suggestion. The detail of a choice of CNN architecture are explained in lines 425-429 page 18. Moreover, we also added the reference of this CNN architecture in [56]. Due to the time limitation, we added your suggestion about optimized the CNN architecture parameters in the conclusion section in lines 558-559 page 25 and we would like to do it for our future work.

Section 4.3 is not explained well, in terms of definition of "image conditions". Is it pose? Illumination? Composition? Color?

Thank you very much, the terms of the definition of "image conditions" are described in  section 3.4. Dataset in line 375-395 page 15-16. We also updated in section 4. Experimental results with adding the details of image conditions in lines 435-438 page 18.

Small punctuation error: line 392.

Thank you so much for your comment. We have removed the comma error at line 414.

Dataset and comparison: only one dataset was used. For future research, I recommend to use, for comparison and better testing of the proposed approaches, datasets from University of Graz: https://www.tugraz.at/institutes/icg/research/team-bischof/learning-recognition-surveillance/downloads/

Thank you very much for value comment. Your suggestion dataset also relates to our image conditions, such as challenging lighting conditions, occlusion, and multiple people for testing. We have added the recommend dataset and other datasets in the references [49-53] related to our image conditions and also mentioned in section 3.4. Dataset in lines 377, 379, 382, 385, 390, and 394. page 15-16. Owing to the time limitation, we have added your suggestion about dataset and comparison in the conclusion section in lines 556-558 page 24-25 and we would like to do it for our future work.

Reviewer 2

Comment:

Response to comment:

The paper presents a lower-body detection framework using proposed anthropometric ratios and compares the performance of deep learning (convolutional neural networks and OpenPose) and traditional detection methods. The paper is very well-structured and justified. The subject is really interesting. Only the authors must cite more recent studies to justify the current applicability of the study.

We very much appreciate reviewer's suggestion, and we have updated the introduction and related work section with addition of recent publications as follows:

Introduction section:

·       Lines 30-33 page 1 for applications in human body detection

·       Lines 43-44 and 48-50 page 2 for human body detection

·       Lines 59-61 page 2 for computer vision for detecting people

·       Lines 85-87 page 2 for anthropometrics application

Related work section:

·       Lines 110-115 and 120-123 page 3-4  for background subtraction methods

·       Lines 127-130 and 144-147 page 4 for object-based methods