On Training Road Surface Classifiers by Data Augmentation

Round 1

Reviewer 1 Report

This work attempts to reduce the captured dataset in size required to train road surface classifiers. 

1. Is this work the first one of “a method to reduce the size of the captured training dataset by augmenting it with synthetic data”? If not, please provide some references to related works and explain what’s novel in this work.
2. In Figure 3, it seems the line “No data augmentation” may grow faster than “Data augmentation”. Can you provide an explanation? Do you expect the two lines may meet with a larger training set?
3. Please provide more explanations of the purpose of the microphones. What is the motivation to use microphones? Maybe some photos of the setup may help.
4. Since some readers are not in the direct research area of road surface classifier, it is recommended the authors give more explanation of what is the training set. The training sets do not include the 10 channels sensors data, right?

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

The manuscript proposes a data augmentation based method to reduce the size of the captured training dataset for road surface classification. The oversampling method GANSO is applied to generate synthetic feature vetors in the training process of the classifier. Experimental results demonstrate that significant saving of signal acquisitions can be obtained by data augmentation. The manuscript is easy to follow. However, some major concerns are as follows:

1. The literature survey is not sufficent. As mentioned in Part 1, there is now a lot of work on road surface classification and road anomaly detection. However, the reference articles [1-5] are too old to reflect the current research trend and advanced nature. In addition, the authors does not compare the state-of-the-art data augmentation based methods for road surface classification. More in-depth literature survey is suggested.
2. The technical contributions are not very clear. Since a combination of the existing methods is used, a clearer context on the technical contributions needs to be provided.
3. In Part 4, the authors does not explain why they choose the Linear Discriminant Analysis (LDA) method for feature ranking. Since there are many kinds of classifiers, further analysis is suggested to demonstrate the practicality of the method.
4. The comparison experiment is insufficent. The experimental results are lack of comparison with other methods and comparison with other SOTA methods needs to be added.
5. There should be a discussion section to state the limitation of the method, as well as discussion of potential improvements.
6. There some grammar issues in the manuscript. For example, in abstract part, ‘It has been demonstrated in the experiments that data 18 augmentation allows a reduction by an approximate factor of 1.5 in the size of the captured training 19 dataset.’, and in Part 4, line 164, ‘The cardinality of the TS will be denominated training set 164 size (TSS).’.

Please check the manuscript carefully.

Comments for author File: Comments.pdf

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

The article deals with the crucial problem for machine learning, which is training classifiers.

In the opinion of the reviewer, minor revision is required.

In section 3 it is not clear what is new in the presented method described in Figure 2 compared to the earlier works of the authors

Section 4 does not describe exactly how the accuracy of the classifiers was determined.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The revised version addressed all my questions. 

Author Response

Thank you for your detailed revision

Reviewer 2 Report

All the concerns have been addressed.

Author Response

Thank you for your detailed revision