Bayesian Linear Regression and Natural Logarithmic Correction for Digital Image-Based Extraction of Linear and Tridimensional Zoometrics in Dromedary Camels
, Francisco Javier Navas González 1,2,*, Elena Ciani 3, MarÃa Esperanza Camacho Vallejo 2 and Juan Vicente Delgado Bermejo 1Round 1
Reviewer 1 Report
The paper proposes a method for extracting camel zoometric data. The paper requires substantial English editing. There are many errors in regards to preposition usage, verb tense, and phrase structure. There is excessive passive voice, which makes the text to be very repetitive. I suggest the authors use it only when necessary.
This is a very specific research topic and I believe the introduction requires a more detailed explanation for broader audiences. First, I believe the authors should start off the introduction by defining zoometry, followed by the importance of zoometry, and how has it been applied so far. After preparing the readers for this topic, explain in more detail why is it so important to do so in camels, and what is the actual research gap in this field. In my opinion, the order of the information displayed does not contribute to an easy understanding of the topic and may lose the interest of the readers.
I even suggest the authors make a “Related Works” section after the introduction to explain in more detail what kinds of attempts have been made so far.
The novelty of this paper is not convincing as it is. Please write the novelties in a clearer way.
A methodological flowchart should be inserted for easing the interpretation of this paper.
This paper has very few visual elements such as figures, which is not a problem. However, for better readability, I suggest the authors list some information in the form of tables rather than plain text, such as the statistics listed in section 2.2.
Section 2.3.1 is very hard to follow. Please try writing this section in a simpler manner.
Section 2.3.2 would be very nice to have some visual elements to show the front, lateral and back views, and if possible, even the images of specific materials that were used and the display of the operators.
The numbers in figure 1 are very hard to see. Please increase them.
Author Response
Reviewer 1
The paper proposes a method for extracting camel zoometric data. The paper requires substantial English editing. There are many errors in regards to preposition usage, verb tense, and phrase structure. There is excessive passive voice, which makes the text to be very repetitive. I suggest the authors use it only when necessary.
Response: The Manuscript was checked by a Cambridge ESOL examinations instructor to improve readability correct verb tense incongruencies and other grammar issues and typos. The use of passive voice was reduced as suggested.
This is a very specific research topic and I believe the introduction requires a more detailed explanation for broader audiences. First, I believe the authors should start off the introduction by defining zoometry, followed by the importance of zoometry, and how has it been applied so far. After preparing the readers for this topic, explain in more detail why is it so important to do so in camels, and what is the actual research gap in this field. In my opinion, the order of the information displayed does not contribute to an easy understanding of the topic and may lose the interest of the readers.
Response: We followed the suggestions by the reviewer and introduction was completely rearranged and completed to include reviewer’s comments.
I even suggest the authors make a “Related Works” section after the introduction to explain in more detail what kinds of attempts have been made so far.
Response: We have included those attempts in the introduction, and further information has been provided in their regards. As the paper is not a review, we do not think it is appropriate to add a related works section.
The novelty of this paper is not convincing as it is. Please write the novelties in a clearer way.
Response: We clarified this in the body text as suggested.
A methodological flowchart should be inserted for easing the interpretation of this paper.
Response: A flowchart was added to make interpretation easier.
This paper has very few visual elements such as figures, which is not a problem. However, for better readability, I suggest the authors list some information in the form of tables rather than plain text, such as the statistics listed in section 2.2.
Response: We added the statistics listed in section 2.2. as Figure 1.
Section 2.3.1 is very hard to follow. Please try writing this section in a simpler manner.
Response: We followed the reviewer’s suggestions.
Section 2.3.2 would be very nice to have some visual elements to show the front, lateral and back views, and if possible, even the images of specific materials that were used and the display of the operators.
Response: We followed the reviewer’s suggestions.
The numbers in figure 1 are very hard to see. Please increase them.
Response: We followed the reviewer’s suggestions.
Reviewer 2 Report
Summary:
In this work, the feasibility to extract 15 camel zoometric data from 2D digital images is evaluated. 30 zoometric measures are collected on field with a non-elastic 17 measuring tape. A scaled reference is used to extract measurement from images.
The manuscript is interesting; however, the following comment need to be addressed carefully :
Title :
- - - - - -
1 – The title is very long; it is suggested to shortened the title .
Abstract:
- - - - - - - - - -
2 – The abstract should be self-contained; therefore, the results should be included in the abstract. The results can be included in terms of improvement ratio between the presented and exiting works .
Introduction Section :
- - - - - - - - - - - - - - - - - - -
3 – Contribution should be included as a list at the end of this section .
Materials and Methods Section :
- - - - - - - - - - - - - - - - - - - - - - - - - - - -
4 – In line 103, the author mentioned “… analysis was performed using the https://scholar.google.com/ search engine …”; it is suggested to rephrase this sentence. “… analysis was performed using Google Scholar search engine …”.
5 – subsection 2.1 need to be rewritten to make it more clear to the readers .
6 – It is recommended to include Table S1 in the manuscript .
7 – Equation in line 133 need to be numbered .
8 – It is suggested to include a table of the data given in section 2.2 .
9 – All Equation should be given a number .
10 – A flow diagram need to be included for the method performed in this work .
Results Section :
- - - - - - - - - - - - - - - - - - -
11 – This section is fine . No comments .
Discussion Section :
- - - - - - - - - - - - - - - - - - -
12 – This section is fine . No comments .
Conclusions Section :
- - - - - - - - - - - - - - - - - - -
13 - This section is fine . No comments .
Genera Comments :
- - - - - - - - - - - - - - - - - - -
14 – There are grammatical error need to checked and corrected . Please check the entire manuscript for grammatical errors and typos.
Author Response
Summary:
In this work, the feasibility to extract 15 camel zoometric data from 2D digital images is evaluated. 30 zoometric measures are collected on field with a non-elastic 17 measuring tape. A scaled reference is used to extract measurement from images.
The manuscript is interesting; however, the following comment need to be addressed carefully :
Response: We thank the reviewer for his/her kind comments.
Title :
- - - - - -
1 – The title is very long; it is suggested to shortened the title .
Response: We followed the reviewer suggestion and title was shortened.
Abstract:
- - - - - - - - - -
2 – The abstract should be self-contained; therefore, the results should be included in the abstract. The results can be included in terms of improvement ratio between the presented and exiting works .
Response: We added our results. Abstract should not contain hints of discussion as there is a specific section for it. We added Cronbach’s alpha results, which is a ratio of scale consistency as suggested..
Introduction Section :
- - - - - - - - - - - - - - - - - - -
3 – Contribution should be included as a list at the end of this section .
Response: We followed the reviewer’s suggestion.
Materials and Methods Section :
- - - - - - - - - - - - - - - - - - - - - - - - - - - -
4 – In line 103, the author mentioned “… analysis was performed using the https://scholar.google.com/ search engine …”; it is suggested to rephrase this sentence. “… analysis was performed using Google Scholar search engine …”.
Response: We followed the reviewer’s suggestion.
5 – subsection 2.1 need to be rewritten to make it more clear to the readers .
Response: We followed the reviewer’s suggestion.
6 – It is recommended to include Table S1 in the manuscript .
Response: We followed the reviewer’s suggestion.
7 – Equation in line 133 need to be numbered .
Response: We followed the reviewer’s suggestion.
8 – It is suggested to include a table of the data given in section 2.2 .
Response: We added a graph as suggested by other reviewer.
9 – All Equation should be given a number .
Response: Equations were all numbered.
10 – A flow diagram need to be included for the method performed in this work .
Response: We followed the reviewer’s suggestion.
Results Section :
- - - - - - - - - - - - - - - - - - -
11 – This section is fine . No comments .
Response: We thank the reviewer for his/her work.
Discussion Section :
- - - - - - - - - - - - - - - - - - -
12 – This section is fine . No comments .
Response: We thank the reviewer for his/her work.
Conclusions Section :
- - - - - - - - - - - - - - - - - - -
13 - This section is fine . No comments .
Response: We thank the reviewer for his/her work.
General Comments :
- - - - - - - - - - - - - - - - - - -
14 – There are grammatical error need to checked and corrected . Please check the entire manuscript for grammatical errors and typos.
Response: The Manuscript was checked by a Cambridge ESOL examinations instructor to improve readability correct verb tense incongruencies and other grammar issues and typos. The use of passive voice was reduced as suggested.
Reviewer 3 Report
This paper studies the feasibility to glean repeatable and reliable camel zoometric data from 2D images, They propose three correction rounds to do so. Here are a few comments that came to mind:
- I am still having a bit of a hard time understanding the main contributions of this paper. could the authors elaborate on that? As of now, it appears that this work is an amalgamation of other methods put together.
- How does this can be generalized for a higher number of data? It appears it was only done for 30 data samples. What is the effect of data size?
- What is the computational cost? how does one do inference? For instance for Bayesain model selection, how would one compute the integral if not tractable?
- Regarding the simulations, I am unsure how one would be able to reproduce their results.
- How does their method compare to other existing methods? How are hyperparameters of the model selected?
Author Response
This paper studies the feasibility to glean repeatable and reliable camel zoometric data from 2D images, They propose three correction rounds to do so. Here are a few comments that came to mind:
Response: We thank the reviewer for his/her work and the attention paid to our manuscript.
- I am still having a bit of a hard time understanding the main contributions of this paper. could the authors elaborate on that? As of now, it appears that this work is an amalgamation of other methods put together.
Response: Digital imaging zoometry has always focused on linear zoometrics and no previous attempt to perimeter extrapolation has been issued, thus tridimensional measurements had been skipped. Furthermore, when inconsistencies or incongruencies existed, no correction method has been addressed. This is the main contribution of this paper, not only because this is the first time done in dromedary camels, but in the rest of species. We clarified this in the body text as suggested.
- How does this can be generalized for a higher number of data? It appears it was only done for 30 data samples. What is the effect of data size?
Response: Special attention was paid to ensuring replicability of results along the manuscript. The sample used is not 30 but 130 animals, which is a remarkably higher number of animals than that reported in previous literature. Furthermore, the whole statistical methods described here followed a process of cross validation and parameters such as BIC, which determines the ability of models to predict for future data to be tested were calculated. Furthermore, the evaluation of Cronbach’s alpha or average measurements of intraclass correlation coefficient determine the validity, accuracy and reliability and repeatability of results. This was clarified in the body text.
- What is the computational cost? how does one do inference? For instance for Bayesain model selection, how would one compute the integral if not tractable?
Response: As suggested in Koehrsen [1], Bayesian linear regression uses probability distributions rather than point estimates. This means, response, y, is not estimated as a single value, but is assumed to be drawn from a probability distribution. The output, y is generated from a normal (Gaussian) Distribution characterized by a mean (the transpose of the weight matrix multiplied by the predictor matrix) and variance (the square of the standard deviation σ, multiplied by the Identity matrix given it is a multi-dimensional model formulation).
The objective of Bayesian Linear Regression is to determine the posterior distribution for the model parameters, rather than the best value for model parameters. Not only the response generates from a probability distribution, but the model parameters presumably come from the same distribution. The posterior probability of the model parameters is conditional on the training inputs and outputs. In contrast to frequentist Ordinary Least Squares Regression (OLS), there is a posterior distribution for the model parameters proportional to the likelihood of the data multiplied by the prior probability of the parameters.
This implies to two primary benefits of Bayesian Linear Regression, priors and posteriors. When there is knowledge, or a guess for what the model parameters should be, these priors can be included in the model (for example the influence of live body weight on zoometrics). This approach contrasts the frequentist approach which assumes everything there is to know about the parameters comes from the data. Indeed, in Bayesian regression, when there is no prior information known, non-informative priors for the parameters such as a normal distribution can be used.
Afterwards, posteriors, or the results of performing Bayesian Linear Regression, which are a distribution of possible model parameters based on the data and the priors. Posteriors enable to quantify the uncertainty about the model. Hence, the fewer data points, the greater the dispersion of posterior distribution will be. As the amount of data points increases, the likelihood washes out the prior, and in the case of infinite data, the outputs for the parameters converge to the values obtained from OLS.
Summarizing, in Bayesian inference for linear regression we use priors as initial estimates, and as we gather more evidence, testing our model against data (posteriors), the model supports or disproves our prior hypotheses. In practice, the evaluation of the posterior distribution for the model parameters is intractable for continuous variables when we implement Bayesian linear regression. Thus, sampling methods are used to draw random samples from the posterior distribution to approximate the posterior distribution to which it should be using Monte Carlo algorithms method and its variants.
- Regarding the simulations, I am unsure how one would be able to reproduce their results.
Response: No simulation was performed. All data were real measurements. The equations issued can be used to replicate results, as reported in the body text. Researchers aiming at replicating results, should go to the field, measure animals, and then apply the equations in this paper to do tridimensional measurement correction, as tridimensional measurements cannot be directly taken from images.
- How does their method compare to other existing methods? How are hyperparameters of the model selected?
Response: Comparison to other methods was discussed. This was clarified in the body text. Hyperparameter selection was clarified in the body text.
Round 2
Reviewer 1 Report
The authors provided the necessary changes.
Author Response
Response: We thank the reviewer for his/her kind comments.
Reviewer 2 Report
In this work, the feasibility to extract 15 camel zoometric data from 2D digital images is evaluated. 30 zoometric measures are collected on field with a non-elastic 17 measuring tape. A scaled reference is used to extract measurement from images.
In the revised manuscript, the authors have addressed the raised comments. No further comments.
Author Response
Response: We thank the reviewer for his/her kind comments.
Reviewer 3 Report
I appreciate the authors' effort to answer my questions.
I understand how Bayesian approaches work, and although I appreciate the authors explaining Bayesian methods, my question was not about how they work. I am still not sure about the computational costs. What is the order of complexity of this approach? is it the same as Bayesian regression? how would it change with increasing number of paramteres? My question is about the computational cost for inference. what is the method? is your model so simple (conjugate prior) that you do not need to employ any approximation inference? it is mentioned MC algorithms, what is the method o choice? Gibbs, MH?
I also understand those are real measurements but you run your simulation (code) to perform an analysis. reproducibility comes down to How could one take your real measurement, apply your method and get the same results as you provided.
Author Response
Comments and Suggestions for Authors
I appreciate the authors' effort to answer my questions.
Response: We thank the reviewer for his/her kind comments.
I understand how Bayesian approaches work, and although I appreciate the authors explaining Bayesian methods, my question was not about how they work. I am still not sure about the computational costs. What is the order of complexity of this approach? is it the same as Bayesian regression? how would it change with increasing number of parameters? My question is about the computational cost for inference. what is the method? is your model so simple (conjugate prior) that you do not need to employ any approximation inference? it is mentioned MC algorithms, what is the method o choice? Gibbs, MH?
Response: We performed Bayesian regression analyses as it is stated in the body text. Hence, the order of complexity is the same. The model is simple, it just regresses one measurement (the one taken on field) against the same measurement (but digitally measured from photos) and the weight (only if explanatory potential and Intra class correlation coefficients sufficiently increase over the levels the ensure solid model consistency as described in the body text).
It would not make sense to think of increasing parameters because the disagreement between a measurement taken on-field and the same measurement taken from digital images is presumably low. Indeed, as shown in the body text, direct translation of measurements reaches acceptable levels of consistency in the most of the cases without the need to perform Bayesian regression. Only in those cases in which an effect of weight on the measurements taken was detected, regression model considered weight as an additional regressor and if necessary, a correction was applied, as it occurred in circumferences and perimeters.
The equations issued in the present paper are meant to be easily and computationally cheap strategies that can be applied straightforwardly. Increasing the number of parameters may directly be against these premises. Approximation was not needed. Jeffreys-Zellner-Siow (JZS) prior was used as a reference prior for the reasons explained in the body text, as it is the most preferable option.
The (Metropolis-Hastings) random walk algorithm which uses Markov Chains to perform Monte Carlo estimate via the Gibbs Sampler algorithm was used as aforementioned given a different prior to the default uniform prior specified in IBM SPSS Statistics Algorithms v. 25.0. by IBM Corp. [59] was selected. The random walk Metropolis algorithm is the preferable option for data imputation from the collection of Markov Chain Mote Carlo (MCMC) as suggested in [60] given neither Admissibility nor Stability is selected.
We clarified this in the body text.
I also understand those are real measurements but you run your simulation (code) to perform an analysis. reproducibility comes down to How could one take your real measurement, apply your method and get the same results as you provided.
Response: As suggested in Bunting, et al. [1], the intraclass correlation coefficient (ICC) is one of the most preferable methods to be used to determine the reproducibility and reliability of numeric measurements organized into groups beyond a simple pairing, for example, different operators measuring the same variable in different animals or the same operator using different methods on different animals.
In this study, we issued the equations and equations were solved. Then, we used ICC to compare the results from model solving and real measurements to test for reproducibility and reliability of models.
We clarified this in the body text.
Round 3
Reviewer 3 Report
Thank you for your response. There is some grammatical error there in the manuscript.
