MtAD-Net: Multi-Threshold Adaptive Decision Net for Unsupervised Synthetic Aperture Radar Ship Instance Segmentation
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsIn this paper, the authors proposed a multi-threshold adaptive decision net for unsupervised SAR ship instance segmentation. The title is interesting, but should be improved in following aspects.
1. There are many spelling and format errors. Taking the abstract as example, in line 6 the authors say "we design a local U shape Feature Extractor(MCTM) to get a threshold vector..." . Instead, it is the "Multiple CFAR threshold-extraction Module(MCTM)". In line 5, the error use of Article"an Multiple-threshold ..." should be "a Multiple-threshold...". Such errors can be found frequently throughout the paper, including misuse of uppercase and lowercase, inappropriate Indentation, and grammar errors, etc.
2. In subsection 3.2, for the function of RemoveLargeValue(), a threshold is set at 250. Why do you choose this threshold? Since different datasets are used in experiments, is this threshold adaptaple to different dataset?
3. In algorithm 2 of subsection 3.2, what's Hist and Pa? The definitions are not given.
4. In subsection 4.2, the experiments are conducted given a fase alarm rate Fa={0.005,0.01,0.1,0.25,0.5}. What if the Fa changes? Different combinations of Fa should be conducted to illustrate the effectiveness of proposal.
5.In experimental section, for the ship segementation results, two kinds of figures are plotted, such as fig.6,8,10,12 for qualitative analysis and fig7, 9, 11,13 for difference illustration. However, these two types are redundant with each other. Odd numbered figures are suggested remained, and efficient for qualitative analysis.
6. In 4.5 Alabtion study subsection, the authors ormited the LUFE and GVTE module because of the PLC-loss computes based on their output. However, these two modules can be replaced by available structures. It is suggested authors add such experiments on different modules to illustrate the effectivness of the proposed structure. The post-processing module "Label smoothing module" should be consider in this part, as well.
7. The output of MtAD-Net will be a threshold dubbed as Tfinal. However, in experimental parts, the computed threshold should be given and analyzed. What's the realtion between the computed adaptive threshold and the given false alarm rate vector?
Comments on the Quality of English Language
There are many grammar and format errors, which should be corrected.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThe authors present an Multi-threshold Adaptive Decision Network dubbed MtAD-Net to segment SAR ship images under unsupervised conditions and have good performance. Here are some comments:
1. In the introduction section, although the contributions of this paper are emphasized, the specific motivations for the proposed method and the specific differences from existing methods are not described.
2. When listing relevant literatures, there is no need to present the main content of each literature in a separate paragraph. It is necessary to describe the associations and essential differences among various methods, rather than simply enumerating the content of the literatures.
3. Figure 3 is not properly positioned and is not centered. The input symbols are too large, and the font of the subheadings is inconsistent with that of the chart titles.
4. In the segmentation results, some are too small while others are too large (as shown in Figure 10 to Figure 13).
5. All the segmentation results shown are the segmentation results of slices. Why not complete the experiments on the test images in SSDD and HRSID?
6. Can the proposed method still effectively distinguish objects from the background in some scenes with complex backgrounds?
7. Although the Transformer is a relatively popular encoding and decoding architecture, the advantages of using it are not explained in the article.
Comments on the Quality of English LanguageThe english writting can be improved.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsThis study focused on unsupervised ship segmentation from SAR images using the proposed method. The approach will be helpful to work with satellite image however, several issues should be addressed before this paper can be further considered for acceptance.
1. The methodology is not clearly mentioned for the processing flow.
2. Each step of Figure 1 should be explained with a logical reason for its necessity.
Comments on the Quality of English Language
Please check some errors.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf