Oriented Object Detection Based on Foreground Feature Enhancement in Remote Sensing Images

Round 1

Reviewer 1 Report

This manuscript proposed a foreground feature enhancement method that can be used for one-stage object detection. The proposed method mainly consists of two key components (keypoint attention module and prototype contrastive learning module). The keypoint attention module is used to enhance the features of the foreground part of the image and weaken the features of the background part of the image. The prototype contrastive learning module is utilized to enhance the discrimination of samples between foreground categories and reduce the confusion of samples between different categories. Besides, the proposed method adopted an equalized modulation focal loss (EMFL) to optimize the training process of the model and increase the loss weight of the foreground later in the model training. The experimental results showed that the proposed method outperforms other one-stage object detection methods. In summary, the content of the manuscript is innovative. And this manuscript is well organized and written. However, I still have some little suggestions for modification.

1)      Please clearly explain the formula (18.a) and (18.b).

2)      In line 372, Figure 8, please indicate the specific expression for the plotted function

3)      Please explain the implications of the bold and underlined data in Table 1.

4)      Please clearly indicate the definition of the mAP in the tables.

5)      Finally, the English expression should be further improved.

Author Response

Dear the Reviewer 1,

We would like to express our sincere gratitude to the two anonymous reviewers for their time and suggestions. We found the two anonymous reviewers’ comments very useful in the improvement of this paper. All the comments have been seriously considered and carefully addressed in the revised manuscript.

Further, we will explain the answers to the comments from the Reviewer 1 one by one in the attached file. The corrected and modified parts in the revised version of the manuscript are marked in RED color, in order to facilitate reading.

At last but not least, we would like to take this opportunity to thank the Reviewer 1 again for his/her insightful comments and valuable suggestions, which greatly helped us to improve the technical quality and the presentation of this manuscript.

Sincerely Yours,

P. Lin, X. Wu, and B. Wang

Dec 2, 2022

Author Response File: Author Response.pdf

Reviewer 2 Report

This paper proposes a single-stage, anchor-free detector with object feature enhancement. The method is described clearly, and the experiment is sufficient and reliable. Suggestions on the revision of the paper are as follows:

 

Major：

1. This paper emphasizes the advantages of the proposed method for fine-grained classification (i.e., similar categories). Therefore, it is recommended to select several groups of category pairs with similar morphology and analyze the classification performance of the proposed method on these categories. For example, the category name can be marked in Figure 11 and analyzed by inter-class and intraclass distances.

 

2. KAM module is similar to MDA-Net in SCRNet [1], but the difference is that this paper proposes the soft label to train the attention module. The authors should demonstrate the advantages of the new method in the ablation study (i.e., soft label with rotation vs. hard label without rotation).

 

3. The motivation of the EMFL module is similar to that of EFL [2]. It is recommended to emphasize the innovation of the EMFL module. In addition, the training details of EMFL should be elaborated, e.g., the rules of the selected samples, how many features are selected for each object? Whether to select only the feature at the object center point according to the bbox annotations?

 

Minor：

1. This paper only uses P2 to participate in prediction. Why doesn't the author try to use P2, P3 and P4 simultaneously?

 

2. In Eq. 18, f_i should be f_ci, and f_k should be f_ck。

 

3. There is no good understanding of why the proposed method has a small performance gain in multi-scale training/testing. It is suggested that the author should explain it more understandably.

 

[1] SCRDet： Towards More Robust Detection for Small, Cluttered and Rotated Objects

[2] Equalized Focal Loss for Dense Long-Tailed Object Detection

Author Response

Dear the Reviewer 2,

We would like to express our sincere gratitude to the two anonymous reviewers for their time and suggestions. We found the two anonymous reviewers’ comments very useful in the improvement of this paper. All the comments have been seriously considered and carefully addressed in the revised manuscript.

Further, we will explain the answers to the comments from the Reviewer 2 one by one in the attached file. The corrected and modified parts in the revised version of the manuscript are marked in BLUE color, in order to facilitate reading.

At last but not least, we would like to take this opportunity to thank the Reviewer 2 again for his/her insightful comments and valuable suggestions, which greatly helped us to improve the technical quality and the presentation of this manuscript.

 

Sincerely Yours,

P. Lin, X. Wu, and B. Wang

Dec. 2, 2022

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

 

The authors have addressed my concerns. It is recommended to accept this manuscript for publication.