Gait Recognition via Deep Learning of the Center-of-Pressure Trajectory

Round 1

Reviewer 1 Report

The manuscript presents an approach for gait recognition using gait signatures extracted using COP trajectory. Deep learning and transfer learning approaches are used for classification. The proposed approach is tested experimentally, and shows good results. Although the results of the proposed approach are good, these are not conclusive because of the small sample size. The accuracy may drop with increase in number of individuals in the dataset. As other methods for gait recognition have used different datasets, the comparison of accuracy is not fully justified. The manuscript presents a novel approach and proof of concept. Overall the manuscript is well written and interesting. The approach may be useful is some practical applications.

Author Response

Thank you for reviewing my manuscript

 

“Although the results of the proposed approach are good, these are not conclusive because of the small sample size.”

I agree that the main weakness of the study is the limited number of participants. However, the number of included strides (108’000) by far exceeds that of previous studies (Table 4). Regarding the uniqueness of COP shapes in the population, I wrote a new paragraph to better explain my views (l 453-464).

“The accuracy may drop with increase in number of individuals in the dataset.”

To address this point, I added a supplementary experiment. Using transfer learning, I tested CNNs on the full database containing 36 individuals (20% larger dataset). The accuracy does not drop at all between 30 and 36 individuals included into the database (see Table 2), which support the hypothesis that COP trajectory is sufficiently unique to constitute a valuable modality for biometric recognition.

 

“As other methods for gait recognition have used different datasets, the comparison of accuracy is not fully justified.”

A new sentence has been added to highlight this issue (line 440-441).

Reviewer 2 Report

A method for “gait recognition via deep learning of the center-of-pressure trajectory” is interesting, and the results appear to be correct, but I found a few technical drawbacks in the manuscript. Therefore, I recommend the minor revision of the paper.

 

The format of Table 2 is error. I suggest the author can apply a set of precision, recall and F1-score measures to evaluate the proposed system. The evaluation about computing power, memory can be listed in this manuscript. A careful English language proof reading would also help.

Comments for author File: Comments.pdf

Author Response

Thank you for reviewing my manuscript

 

“The format of Table 2 is error.”

The Table 2 has been modified to include new results.

 

“I suggest the author can apply a set of precision, recall and F1-score measures to evaluate the proposed system. “

First, I considered to include other accuracy indexes in addition to overall accuracy, such as confusion matrices and F1-scores per class. However, the accuracy I obtained was so high that reporting them was not informative. With only 3-4 misclassified items, the F1-scores was equal to 1 for most classes, and then the average F1 score was virtually equal to 1. Given that the class were balanced (i.e., each class is represented by an equal number of items), using the accuracy (rate of correctly classified items) seemed a better option. In my opinion, overall accuracy gives a more direct feeling about the correctness of classification. In contrast, for the verification experiment, I used AUC and EER which are more appropriate for unbalanced classes. I slightly modified the explanation about EER and AUC in the revised manuscript to better explain this choice (lines 321-324).

 

“The evaluation about computing power, memory can be listed in this manuscript.”

All the computing can be reproduced on CodeOcean. The link is in the beginning of the manuscript. I added some indications about the current cloud computing environment and about the duration of computation (Line 353).

 

 “A careful English language proof reading would also help. “

 

The revised manuscript has been edited by a professional service.

Reviewer 3 Report

The paper presents a method for individual identification using center of pressure using convolutional neural nets. Overall the paper is well written.  The main concern I have is with the limited scope of the data set, and problems with overfitting and model brittleness.  The authors do state these limitations in the discussion, but addressing them in the study is preferable.  If the authors collect some additional data on a small set of subjects under different conditions, speed, shoes, slope, etc.  The effectiveness of the model then can be more fully tested.  This will greatly improve the impact of the paper without requiring a large recollection of data.  

Author Response

Thank you for reviewing my manuscript

 

“The main concern I have is with the limited scope of the data set”

I added a new paragraph to better highlight this issue (l 453-464).

 

,” and problems with overfitting and model brittleness. “

 

In my opinion, enough caution has been taken to avoid overfitting:  separated sets for parameter tuning and for validation, L2 regularization, batch normalization, and early stopping. Overfitting in deep learning is a complex question: recent advances have shown that over-parametring NNs is rather beneficial and that the classical views about overfitting are not valid for deep NN, see for instance: https://arxiv.org/pdf/1901.01608.pdf. What other steps would you suggest taking to further reduce the risk of overfitting?

 

“The authors do state these limitations in the discussion, but addressing them in the study is preferable.  If the authors collect some additional data on a small set of subjects under different conditions, speed, shoes, slope, etc.  The effectiveness of the model then can be more fully tested.  This will greatly improve the impact of the paper without requiring a large recollection of data. “

I agree that further experiments are need to validate the method more thoroughly. However, I have no possibility to collect new gait data. The dataset was collected few years ago, and currently I have no access to an instrumented treadmill. Actually, three different walking conditions were used in the present study (see lines 528-538).

I hope other researchers can replicate my results and extend them to other walking conditions. I modified the conclusion to reflect this point of view (lines 615-616).

 

Round 2

Reviewer 3 Report

The paper presents a method for individual identification using center of pressure using convolutional neural nets. Overall the paper is well written.  

I believe the authors have sufficiently highlighted the limitations of the study.  I would agree with the author that further study is needed and hope they again get access to an instrumented treadmill.