Object Detection with Low Capacity GPU Systems Using Improved Faster R-CNN

Round 1

Reviewer 1 Report

This paper introduces a novel method, an improved faster R-CNN, for object detection in remote sensing images. Deformable convolution and FPN have been incorporated to build the multi-scale faster R-CNN. I have the following comments.

The illustration figures need improvement. In figure 2, the meanings of the two arrows are not clear. It's also better to include R (used in the text description) in the figure. The formulae need improvement. In (1) and (2), w and P_o should be defined. In (3), \mu, \sigma and \epsilon should be defined. In the experiments, the random training/test split should repeat several times and record the means and variances as more reliable results. Boxplots can be used to show the results. There's no need to show the F1 scores by bar plots in figures 8, 9 and 10. Show the numbers in one table with largest number in bold is sufficient.

Author Response

Dear Reviewer,

      Thank you for comments concerning our manuscript entitled “Object Detection with Low Capacity GPU Systems Using Improved Faster R-CNN”. Those comments are all valuable and very helpful for revising and improving our manuscript, as well as the important guiding significance to our studies. We have studied comments carefully and have done all corrections. Reponses to comments have been added as attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

General comments:

This manuscript deals with the object detection using a faster R-CNN model. In general, the objectives of this paper are clear, but the analyses are insufficient. Therefore, revisions are required. The following summarize these points in detail.

Major Comments:

What are your real contributions to the Faster R-CNN model? Deformable Convolutional Network, Feature Pyramid Network, and Weight Standardization strategies as you mentioned all have been used by others to improve the CNN models. Have you improved one(or all) of them in this study? What is your new concept through the comparison with others? Or you just simply integrated them to the Faster R-CNN model that no one did before? Please clarify this and reorganize your paper.

Minor Comments:

The full name of every abbreviation should be given at where it first shown. In addition, it is better to list all the abbreviations and their corresponding full names in a table at the end of this paper. Figure 1, please clearly point out the Regional Proposal Network and Object Detecting  Section 2.5 should be moved to the introduction part. It is hard to see the labels in figure 11. The conclusion is insufficient to reflect the authors’ work. Please rewritten this part. Please check the whole paper for typos, grammar errors. At last, this paper is not well organized and the analysis in many places is insufficient.

Author Response

Dear Reviewer,

      Thank you for comments concerning our manuscript entitled “Object Detection with Low Capacity GPU Systems Using Improved Faster R-CNN”. Those comments are all valuable and very helpful for revising and improving our manuscript, as well as the important guiding significance to our studies. We have studied comments carefully and have done all corrections. Reponses to comments have been added as attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Some of my comments are well answered. I till have the following comments.

1. the explanations to the equations still need improvement. For example, in (2) you wrote P_o while in the explanations you wrote p_0. In (3), the font styles of the symbols are different in the equation and in the explanations. You need to make sure that the symbols used are consistent.

2. it is not clear whether the plots in figure 8 are average versions of precision-recall plots for 10 training/test splits or they are one result picked from 10 training/test splits? If is the average version, how did you do the average?

 

Author Response

Dear Reviewer,

Thank you for comments concerning our manuscript entitled “Object Detection with Low Capacity GPU Systems Using Improved Faster R-CNN”. Those comments are all valuable and very helpful for revising and improving our manuscript, as well as the important guiding significance to our studies. We have studied comments carefully and have done all corrections. Reponses to comments have been added as attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Thanks for the authors' reply. I have no further comments.

Author Response

Many thanks for your interest and positive evaluations.