Visual Place Recognition of Robots via Global Features of Scan-Context Descriptors with Dictionary-Based Coding

Round 1

Reviewer 1 Report

Place recognition is essential for robots, and the topic is valuable for the readers. In this manuscript, the authors claimed that they suggest a method called BFG-system to achieve more accurate and faster place recognition. However, the motivation to conduct this study is not presented clearly, and how global features are obtained is not explained in detail. Meanwhile, the study seems just put different algorithms together, while lacks significant contributions. Therefore, I would suggest the authors to extensively improve their manuscript before submission.

1. first of all, what are the special features of the proposed methods, and how the proposed method outperforms existing methods, please provide some details in abstract.

2. what are the problems of scan-context descriptor and is there any other method for place recognition? please provide some evidences or cite some articles to support your claims made in line 47-52.

3. considering the above comment, the motivation to conduct this research is not clearly explained.

4. line 64-70 are not the contributions of a research, please highlight the scientific contributions or improvements made.

5. first paragraph of section 2 should be removed.

6. line 119-124, please provide evidences to support your claims here.

7. review of related literature is too narrow and lacks many investigations related to place recognition.

8. the approach presented is too normal and it is hard to judge the novelty of this method.

9. what are the special improvements and how the global features are obtained are not clearly presented. 

10. line 255-256 is confusing.

overall, the language of this manuscript needs some improvements to present the content more clearly. At it current status, there still exists some grammatical errors or typos. For example, the abbreviation of BGF looks strange.

Author Response

Dear reviewer,

I am honored to receive your review.

Plese see the attachement.

best regards,

Minying YE

Author Response File: Author Response.docx

Reviewer 2 Report

Improved visual robot place recognition of scan-context descriptors by global feature reclassification. (applsci-2394378) 

 

Major Concerns

The motivation, the related works, and the strong points of the presented place recognition method are carefully written in the submitted manuscript. I have NO major concerns about this work itself. However the paper structure and writing issues could be revised as follows.

 

Minor Concerns

[1] Explanation of the methodology

It was not explained sufficiently in the current manuscript. Please explain how to train VGG19 in the Feature extraction phase. The input is a scan-context image… but what is the target label?  Is it the location (the area sets) based on the GPS? 

Also, the training loss of VGG16 in the estimation phase will be not clear to readers. Line 178 says ‘prediction the test datasets’, but which value in the test sets is to be estimated?  Please explain in detail in Section 3.

 

[2] Paper title

I think the dictionary-based coding will be one of key ideas of the presented system and can appear in the title as, for example, “Visual place recognition of robots via global features of scan-context descriptors with dictionary-based coding.” Please reconsider the paper title if possible.

 

[3] Figure 1

The visualized input data are ‘scan-context images’ but they are labeled as ‘Point Cloud’.  I understand that the label ‘Point cloud 2012-01-15’ can be ‘Training SCI set’ and ‘Point cloud 2012-01-08, …,-09-28’ in the bottom part can be ‘Test SCI sets.’

On the top part, ‘K-means and Dimensionality reduction’ will be misleading.  I understand that the dimensionality reduction is applied using a random sampling before the K-means clustering, and actually the K-means is used for the reclassification of the global features. I think the dimension reduction is not important in the presented framework. Thus, the labels on the top part could be ‘Global feature extractor’, ‘Global features’, and ‘Reclassification by a K-means’. 

It would be better to place the ‘dictionary coding’ label somewhere.

‘Based on global feature re-classification place recognition framework’ will be removed.

‘CNN models’ will be ‘Place recognition networks’.

The caption can be polished as: ‘Figure 1. The pipeline of place recognition method using the proposed global features (BGF) based on the scan-context image (SCI)’ for example.

 

[4] Expected effect of the presented method.

In the 3rd paragraph (Lines 47-51) in Section 1, the authors point-out two issues of the previous methods, i.e., data on the boundaries of each region is difficult to predict, and also a heterogeneous distribution of descriptors can be erroneous. However, in the 4th paragraph, the authors have mentioned only the latter issue. I would recommend discussing the former issue in the 4th paragraph of Introduction, and also in Discussion.

 

[5] Explanation of the technical term

I recommend providing a brief explanation of the key jargon, ‘scan-context image’ in Introduction. I understand the scan-context image is a 2D data that reflects the geometry of 3D point cloud where the color shows the distance between the robot and objects.

 

[6] The structure of paper

Section 4 and Section 5 can be a single Section like

4. Experiments

4.1. Datasets

4.2. Baseline methods

4.3. Experimental results

(1) Comparison of accuracy

(2) Global feature analysis

(3) Runtime evaluation

 

[7] The reference for the K-means

I think the paper[20] entitled “Enhanced binary genetic algorithm …” is not a good reference for the K-means.  How about citing “Lloyd, S., 1982. Least squares quantization in PCM. IEEE transactions on information theory, 28(2), pp.129-137.” as [20]?  (This paper is not my work nor my colleagues’.)

 

Typo:

Line 26: research(es) (are) based 

Line 42: Although(,)

Line 47: the original descriptors classification method is mainly divided based on the geographical location. (“method is divided” sounds unnatural. The sentence could be corrected as) –> the conventional methods divide the original descriptors based on the geographical location. 

Line 55: continue to use … and integrate(d) it into our system  –> leverage … in our system.

Line 56: The basic idea was -> Our idea is. (“The basic idea was” sounds like the idea was of the previous study.)

Line 101: and is 2D feature generated (I could not understand this sentence at all.)

Line 119: Although many research(es have focused) on

 

Best regards,

n/a

Author Response

Dear reviewer,

 

Thanks for your comments.

I have responded your comments.

Please check the attachment.

 

best regards,

Minying YE

Author Response File: Author Response.docx

Reviewer 3 Report

The manuscript entitled “Improved visual robot place recognition of scan-context descriptors by global feature reclassification” reclassified the image according to the CNN global features through image feature extraction. The work is interesting and useful, and improved solutions can be found in this work. However, the whole manuscript should be significantly improved before further consideration. The comments are:

1. In abstract, the statement “The results show that Our method is…” should be modified. “Our method” should be changed to “the proposed method”.

2. Figure 2 should be improved as the resoluation is low.

3. To demonstrate the adavantages of the proposed method, more existing algorithms should be taken into account.

4. Although the writing is acceptable, the whole paper should be furhter polished.

5. The manuscript is too slim in terms of an academic paper. It is not a short communication, and therefore more contents should be added, for example, more explanations and figures.

The reviewer will give more detailed comments once the work is improved based on the comments.

Should be further improved.

Author Response

Dear reviewer,

Thanks for your comments.

I have responded your comments.

Please check the attachment.

best regards,

Minying YE

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

The Reviewer is not satisfied with the improvement and Response to Reviewer.

Therefore, the Reviewer has to reject the paper.

The writing can be improved.

Author Response

Dear reviewer,

I apologize very much for not including the new method for comparison experiment. Because I need a lot of time to carry out the compared experiment. So I am so late for uploading the new version. I have added the other paper methods for comparison with ours. I hope you can review it. And highlighting has been done in the parts that have been modified.

 

Best regards,

Minying YE

Author Response File: Author Response.pdf