A Novel Anchor-Free Method Based on FCOS + ATSS for Ship Detection in SAR Images
Round 1
Reviewer 1 Report
- The Abstract should clarify the evaluation criteria of the proposed model and provide the detail of state of art as mentioned in the abstract (in number).
- There are few grammatical and English languages, spelling errors, grammar and punctuation errors are found.
- Please include a discussion about future work in the conclusion.
- What is the limitation of the proposed algorithm?
- Include the most related references, such as:
- Detection and classification of multiple power quality disturbances in Microgrid network using probabilistic based intelligent classifier.
- Multi-objective evolutionary algorithm for pet image reconstruction: Concept.
- Zhang, T., & Zhang, X. (2019). High-speed ship detection in SAR images based on a grid convolutional neural network. Remote Sensing, 11(10), 1206.
- Guo, H., Yang, X., Wang, N., & Gao, X. (2021). A CenterNet++ model for ship detection in SAR images. Pattern Recognition, 112, 107787.
Author Response
Point 1: The Abstract should clarify the evaluation criteria of the proposed model and provide the detail of state of art as mentioned in the abstract (in number).
Response 1: Considering your suggestion, we have re-writted the abstract. (in red)
Point 2: There are few grammatical and English languages, spelling errors, grammar and punctuation errors are found.
Response 2: We are very sorry for these errors. We have checked the full text and made corrections. (See the underlined content)
Point 3: Please include a discussion about future work in the conclusion.
Response 3: We have added a discussion about future work in the conclusion. (in red)
Point 4: What is the limitation of the proposed algorithm?
Response 4: Although the proposed method has better detection accuracy, its detection speed is not the fastest, which requires further analysis and research. (in red)
Point 5: Include the most related references, such as:
Detection and classification of multiple power quality disturbances in Microgrid network using probabilistic based intelligent classifier.
Multi-objective evolutionary algorithm for pet image reconstruction: Concept.
Zhang, T., & Zhang, X. (2019). High-speed ship detection in SAR images based on a grid convolutional neural network. Remote Sensing, 11(10), 1206.
Guo, H., Yang, X., Wang, N., & Gao, X. (2021). A CenterNet++ model for ship detection in SAR images. Pattern Recognition, 112, 107787.
Response 5: Considering your suggestion, we have added these references[10-11] [26-27].
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The paper proposes using group convolutions and detection heads with a regressed label to
improve FCOS+ATTS network. The proposed approach improved AP50 by 4.7%
compared with the baseline without these two ideas.
- The paper lacks details about the training procedure. Please describe it in more detail.
- The authors need to describe how they tuned the parameters. Have you used a validation
set?
- The letters font size in the Equations is tiny compared to the rest of the paper. I suggest
the authors change it to adequate font size.
- Lack of a significance test.
Overall is a good paper but needs some adjusting.
The paper proposes two ideas:
- using group convolutions and
- detection heads with a regressed label to improve FCOS+ATTS network.
The proposed approach improved AP50 by 4.7% compared with the baseline without these
two ideas.
Overall, the paper is well structured but lacks more details for reproducibility and more indepth experimental analysis details. My major suggestions are:
- The paper lacks details about the training procedure. Please describe it in more detail.
- The authors need to describe how they tuned the parameters. Have you used a validation
set?
- Lack of a significance test.
The fulfillment of these items would change the paper substantially, especially if there is no
significant difference in the results considering the proposed approach.
Minor suggestion
- The letters font size in the Equations is tiny compared to the rest of the paper. I suggest
the authors change it to adequate font size.
Comments for author File: Comments.pdf
Author Response
Point 1: The paper lacks details about the training procedure. Please describe it in more detail.
Response 1: Considering your suggestion, we have added the details about the training procedure in Section 3.2. (in red)
Point 2: The authors need to describe how they tuned the parameters. Have you used a validation set?
Response 2: The proposed method is trained with GPU and finished in 30th epochs. The momentum and weight decay are set to 0.9 and 0.0001, respectively. We choose SGD with the initial learning rate of 0.005 as the optimizer, the other hyperparameters are set to the default values in MMDetection. We used a validation set. Half of the images in the training set are used for training and the other half for validation. We have made additional explanations in the text. (in red)
Point 3: The letters font size in the Equations is tiny compared to the rest of the paper. I suggest the authors change it to adequate font size.
Response 3: Considering your suggestion, we have revised these equations.
Point 4: Lack of a significance test.
Response 4: We have added the visual detection results of FCOS+ATSS and the proposed method (in Figure 7). It can be seen that the proposed method can reduce missing detections and false alarms, indicating that the proposed method can improve the detection performance of FCOS+ATSS.
In addition, we have added the visual detection results of different methods (in Figure 8). In the first column of Figure 8, one false alarm exists in other methods, but the proposed method can avoid this false alarm. In the second and third columns of Figure 8, there exist some missing ships in all methods. However, Faster RCNN and our method have fewer missed ships than other methods. In the fourth column of Figure 8, most of the ships were missed by other methods. However, our method has the fewest missing detections. Com-pared with other methods, the proposed method obtains a better detection performance.
Author Response File: Author Response.pdf
Reviewer 3 Report
I understand Formula 1 very little. You have to explain it better.
Figure 1, if you have to insert it, you have to explain it well. You have to explain better what each computational block is used for, then you have to insert the dimensionality of each input and output image map.
Also formulas 2 and 3 need to be explained better.
Figure 4 needs to be explained better.
Figure 5 you need to explain it better and you need to include the dimensionality of the patches.
Author Response
Point 1: I understand Formula 1 very little. You have to explain it better.
Response 1: Considering your suggestion, we have reinterpreted Formula 1. (in red)
Point 2: Figure 1, if you have to insert it, you have to explain it well. You have to explain better what each computational block is used for, then you have to insert the dimensionality of each input and output image map.
Response 2: Considering your suggestion, we have reinterpreted Figure 1. (in red)
Point 3: Also formulas 2 and 3 need to be explained better.
Response 3: Considering your suggestion, we have reinterpreted Formula 2 and 3. (in red)
Point 4: Figure 4 needs to be explained better.
Response 4: Considering your suggestion, we have reinterpreted Figure 4. (in red)
Point 5: Figure 5 you need to explain it better and you need to include the dimensionality of the patches.
Response 5: Considering your suggestion, we have reinterpreted Figures 5. (in red)
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
The authors address the main issues, but the paper still lacks a statistical significance test.
Author Response
Point 1: The authors address the main issues, but the paper still lacks a statistical significance test.
Response 1: Significance test is a method used to detect whether there is a difference between the experimental group and the control group in a scientific experiment and whether the difference is significant. On the one hand, existing literature hardly applies significance test to object detection. On the other hand, object detection usually verifies the effectiveness of the method by comparing whether there is a significant difference between the validation set and the test set. The detection results on the validation set are shown in the figure below. There is no significant difference between the validation set and the test set, which verifies the effectiveness of the method in this paper.
Author Response File: 
Author Response.docx
Reviewer 3 Report
Accepted.
Author Response
Thank you for your suggestions.
