ResiDualGAN: Resize-Residual DualGAN for Cross-Domain Remote Sensing Images Semantic Segmentation
Round 1
Reviewer 1 Report
The paper entitled 'ResiDualGAN: Resize-Residual DualGAN for Cross-Domain Remote Sensing Images Semantic Segmentation' has been reviewed. The paper proposes an unsupervised domain adaptation model named ResiDualGAN to complete cross-domain semantic segmentation tasks. The authors have presented a remarkably adequate description of the model and thorough analysis of the experimental results, which shows a superiority over other methods. It is a topic of interest to researchers in related fields, but the paper still needs some minor revision before it can be accepted. The detailed comments are as follows:
1) From the visual effect, after the image translation of stage A, the translating images from S to T have obvious blurring problems and the presence of shadows causes changes in the geographical features after the translation, will this have a negative impact on the subsequent semantic segmentation task?
2) The image size is strictly limited according to the resolution of the dataset for the purpose of reducing the scale discrepancy, does this lead to the problem of poor generalization of the proposed model and does the resize module still work better if the same image size is chosen when making the dataset.
3) Some parts of the paper contain writing errors and layout design problems.
Author Response
Thanks for your comments. Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The authors developed a resize and residual GAN model for image transforming between different domains obtained by different datasets. The study gave a comparison for translations between PotsdamIRRG, PotsdamRGB, Vaihingen, and BC403. And the segmentation was performed on these pictures. An improved result can be obtained by output space adaptation (OSA). I would like to give my comments as below,
1. It seems that the role and function of OSA was not described in details. Especially the reason why the OSA can improve the accuracy of segmentation was not described clearly in Sec. 4.4. If the authors have an easier way to explain why OSA is superior than feature extraction for studies in remote sensing, it is encouraged to put some description in Sec. 4.4.
2. The use of t-sne analysis needs more detailed information. What are the settings for Fig. 8. Is there a quantitative estimation for the matching between the distributed feature points? A number indicating the degree of matching is better than visualization only.
3. The grammar and the expressions are advised to be checked again though most parts of the manuscript are correct and smooth. There are two verbs in the second sentence in Table 5 for the description of ablation study. The expressions in the paragraph from line 305 to line 308 seem to be not smooth. Revision is advised if necessary.
4. I have a question in Sec. 3.3 on the use of IoU as the index of accuracy in the study of remote sensing. As I understood, IoU might concern more about the similarity of the shapes between the predicted object and the ground truth. In remote sensing, does it matter to pursue the accuracy of shapes of the objects? What about the accuracy of the locations of objects? If the authors have additional ways of accuracy estimation, I would like to encourage them to give comments on the accuracy of their study.
Author Response
Thanks for your comments. Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Remote sensing deals with a plethora of domains. It is not possible to collect annotated data over each domain. Towards this, the manuscript states domain to domain translation as a solution, especially in the context of semantic segmentation. Furthermore, the manuscript proposes two improvements over the existing GAN-based image translation methods. The manuscript needs clarification/improvement in several sections:
1) Much of the proposed method, e.g., Section 2.1.3, 2.1.4, 2.2.2, 2.2.3 are well known information and is not novel to this manuscript. Please emphasize on your novelty.
2) GANs and residual models have been explored jointly in the literature many times. Pleae clarify how your propose scheme is novel.
3) Towards the challenge described in the manuscript (lack of annotated data), several unsupervised segmentation methods have been proposed in the literature. Please clarify how is the proposed method superior to them. Here are a few examples:
i) Hierarchical Object-Focused and Grid-Based Deep Unsupervised Segmentation Method for High-Resolution Remote Sensing Images
ii) Unsupervised single-scene semantic segmentation for Earth observation
iii) Weakly Supervised Semantic Segmentation in Aerial Imagery via Explicit Pixel-Level Constraints
4) Data distribution of Potsdam and Vaihingen is not significant, two cities are fairly similar and located in similar geographic area. I think experiment on them is not sufficient in the context of the problem of this manuscript.
5) Please explain output space adaptation in more details.
Author Response
Thanks for your comments. Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 3 Report
1) Necessity of resizer module is not clear. It serves the scale discrepancy of two domains that could have been addressed by simply applying superresolution to one of the domains.
2) In continuation of above point, please provide comparison where one domain is simply superresolved.
3) Regarding the papers in previous comment 3- even if they are directly not comparable, they must still be included in the literature review.
4) In continuation of above point, authors mainly contended that the described papers are for unsupervised segmentation whereas their work is for domain adpatation, so they are not comparable. Here I do not agree, because if unsupervised segmentation on the target domain unlabeled data already works work, there is no requirement for domain adaptation.
5) In reference to previous comment 4, please include Zurich dataset in the experiments.
Author Response
Thanks for your comments. Please see the attachment.
Author Response File: Author Response.pdf
