Review Reports

Round 1

Reviewer 1 Report (Previous Reviewer 2)

Although the segmentation ability has reached a high level, I can't feel your highlights. You seem to have made two improvements. But in fact, from your paper, I think you just used GDConv instead of Conv in the trunk. I don't see the advantages of GDConv over fixed offset GDConv. You should use these two convolutions to do a comparative experiment, which is far more convincing than language description.

Reviewer 2 Report (Previous Reviewer 1)

The authors address the points discussed in the revisions.

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.

Round 1

Reviewer 1 Report

The manuscript presents a methodology where Gaussian Dynamic Convolution for semantic segmentation in remote sensing images; however it has some flaws:

Given the motivation of the manuscript, the authors claim that the change in scales of the instances of a given category give rise to performance degradation of any method. Then, it is given that variable convolution can alleviate this problem along with irregular shapes of the objects, referenced by the papers [6] and [7]. Within this context, authors present Gaussian Dynamic Convolution. 

However, the experiments and the discussions on the results of the experiments remain quite naive in order to verify the contribution of the paper. The followings are necessary:

1) If any of the methods, Deeplab v3+, RefineNet, PSPNet, FarSeg, FactSeg, involve dilated convolution and deformable convolution, it is required to thoroughly discuss the results.

2) If these methods do not have these convolutions (dilated, deformable), it is expected to perform experiments with the methods using these convolution techniques discussing the results from the convolution technique point of view, since the manuscripts' motivation is based on these.

3) Why are the numbers on the TC, LV, SBF categories lower? Please discuss these points? Does it arise from lack of training data or your convolution technique is no good?

4) There are not any sample for false positives. Please discuss the false positive hypotheses from the convolution technique point of view.

Several things relatively minor:

* It is necessary to give the manuscript a linguistic check. e.g. (Page 1) An unique semantic label, (Page6) ...instances.Since....

* What are ST, BD, TC, BC, ...?

 

Reviewer 2 Report

1. The paper does not have a network structure diagram, so I cannot quickly see the innovation points.

2. Inadequate innovation of the thesis. Simply add modules and perform convolutional replacement. The paper needs its own innovation.