An Illumination Insensitive Descriptor Combining the CSLBP Features for Street View Images in Augmented Reality: Experimental Studies

Round 1

Reviewer 1 Report

The manuscript proposes an improvement to the existing feature descriptors by combining them with the center-symmetric local binary pattern (CS-LBP) to improve illumination sensitivity. The matching experiments were performed on five street view image sequences with different illumination changes. The results showed improved matching accuracy for street view images which have illumination changes with only a little increase of time consumption.

Although the manuscript seems to presents a valuable contribution backed up with significant series of experiments, some issues need to be addressed before publishing this work:

1) [Line 10] The first sentence of the abstract is too long, so I advise rewriting it. The second sentence can also be split into two. Try using more concise sentences.

2) [Line 93] Please consider to rename steps from "StepN" to "Step N" or "Step-N" to increase readability.

3) [Figure 2] The part of the image "Descriptor is binary?" is vague. As I understood, Step 4 is binarization but it is optional. Your block in the flow chart should be something like "If BRISK, ORB, FREAK, or AKAZE?" and then in the YES branch add "Construct binary desc...", while in NO branch put "Normalize the descriptor".

4) [Line 135] "Combine the original description vectors" should be further explained. Are CS-LBP descriptor concatenated to the original descriptor?

5) [Figure 3] I suggest replacing these 8 flow charts that are essentially the same (except the distance block) with one generic that is applied to all 8 original descriptors. Then describe specifics in the image caption.

6) [Table 1-10] The content of the table is not clear. Please change the table content and the caption to better describe present the results. In particular, the first column specifies the original descriptor, but in the second column, we have two rows: CSLBP and <DESC>-CSLBP. The second is the proposed combined descriptor, but what is CSLBP? The table caption says: "The difference values of recall between the CSLBP algorithms and the original algorithms...". What are the values of the original algorithms?

7) [Table 11] The caption says: "The radios of...". I assume you meant "ratios". Again the same comment applies to the content as before.

8) The manuscript should be revised so that a certain part of the appendices should be moved in the main part of the manuscript so it would be easier to follow. Appendix A presents an overview of the 8 original descriptors, and although it is informative I think it is excessive. Appendices B and C should be integrated into the main text. I suggest making a new section that describes used datasets that should contain the content of Appendix B and integrating Appendix C into the Experimental Analyses section. Maybe you can also introduce the Related Work section after the Introduction and combine some parts of the introduction and Appendix A (the figures or the text with steps).

Author Response

Thanks for the reviewer's comments and suggestions, these comments are really helpful for our revision. We looked through your comments and have modified our paper correspondingly. The reply to your review please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

It should be necessary some improvements in English. Many expressions are difficult to understand. The abstract should be summarized and shortened.

Introduction should explain better the goal of the research, and we need sme conclusions explaining practical applications of the research.

Research is interesting and well focused, and explains very well the idea under a technical approach. It should be necessary explain some practical applications of the research in a section where conclusions are described.

Author Response

Thanks for the reviewer's comments and suggestions, these comments are really helpful for our revision. We looked through your comments and have modified our paper correspondingly. The reply to your review please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

The authors present an illumination insensitive matching algorithm for street view images in Augmented Reality integrating the center-symmetric local binary pattern (CS-LBP) into the common feature description framework.

In this way they improve state-of-the-art feature matching algorithms (e.g., SIFT, SURF, DAISY, BRISK, ORB, FREAK, KAZE, and AKAZE) in terms of precision.

The paper is well structured, methods are adequadelty described and experiments show results that support conclusions. Despite a positive opinion on the authors' work, there are some points that can be improved/corrected, in particular from a readability point of view.

The choice of use appendixes leads to difficulties on immediate understanding of the text, in particular while reading experiments description.

The figures with flow charts are too small and they are not easily readable.

The tables with difference values are not easily readable. You could divide CSLBP values from algorithm-CSLBP to put in evidence improvements of your work from original matching algorithm or at least you could put an horizontal line to separate each row couple.

I suggest to extend discussion on result, in particular to comment the general improvement on precision with you algorithm, while the recall in general is lower.

Line 12, 73, 85, 344: personally I prefer other terms instead of obvious/obviously.

From line 74 to line 78: instead of generic word "Reference" is better to use an authors-year style to have, directly on the text, an idea of the research evolution on that field.

Line 202, "#correspondece" should be "#correspondence".

Author Response

Thanks for the reviewer's comments and suggestions, these comments are really helpful for our revision. We looked through your comments and have modified our paper correspondingly. The reply to your review please see the attachment.

Author Response File: Author Response.docx

Reviewer 4 Report

The authors present a combination of approaches, thoroughly tested them against a number of matching cases and present the results in great detail. The experimental results will likely be useful for researchers considering feature descriptors in the future.

However, the text of the paper is not yet in a state ready for publication. First, there are many issues with language and style. Especially in the beginning the majority of sentences need to be improved. I marked the first few occurrences of the more glaring mistakes in the text, but this is not exhaustive – a full pass of editing on the whole text seems necessary.

The paper is also very heavy on repetition. Many sentences are reiterated 8 or more times for each individual descriptor - this is excessive. I would suggest going through all listings and repetitions and consider for every one of them whether it is actually needed – your results are relatively straight-forward, and the writing style makes them needlessly difficult to understand. This especially goes for the result section, consider letting your tables do most of the numbers work and really try to condense the text down to something more palatable and immediately useful for interested researchers – perhaps try to make the conclusions 1 and 2 more like conclusion 3. Only reiterate numbers in the text where especially interesting or necessary.

The text also isn’t clear enough on what the actual contribution is. The CS-LBP descriptor as well as all the other descriptors are from previous publications, and your way of combining them is pretty straight-forward, so it seems like the experiments are the main contribution of this paper – this should be clear both in the introduction and title. (On the same note, I am also not sure about Appendix A. It is surely interesting for readers not familiar with the common descriptors, but it also seems to reiterate work that is already cited in the beginning. I would suggest either removing it or adding some additional value to it, like about performance specifically in AR scenarios.)

In the same vein, the paper needs to be more specific about its use case. AR features very prominently in the abstract and introduction but is not heard from again in the main text. Complex illumination challenges are not exclusive to AR, so this does not work as a motivation on its own. I would suggest either removing the references to AR altogether, or to really consider all the descriptors and their results through the lens of AR – which conditions are common, which descriptors see the most use in current AR, etc.

The paper also seems to be missing references to geo-information. In its current state it would feel more at home in a dedicated computer vision journal.

Please note that while this is a long list of issues, none of them are related to your results or methodology, which both seem solid. If you rework the text as described, this could be a very interesting article.

Comments for author File: Comments.pdf

Author Response

Thanks for the reviewer's comments and suggestions, these comments are really helpful for our revision. We looked through your comments and have modified our paper correspondingly. The reply to your review please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

I am satisfied with how the authors addressed my comments and I advise accepting the manuscript for publication.

Reviewer 4 Report

now it is ok for me