Learning Lightweight and Superior Detectors with Feature Distillation for Onboard Remote Sensing Object Detection

Round 1

Reviewer 1 Report

This paper present a novel method for knowledge distillation for object detection. The methods proposed are shown effective in transferring knowledge from one or more teacher networks to a student network. The proposed methods are adequately compared with other state-of-the-art approaches, and prove to be advantageous for remote sensing tasks. 

The paper is well written and readable. A few comments:

- If the proposed technique is effective for remote sensing object detection, like demonstrated, it would be interesting to see how well it performs on more general object detection datasets (such as e.g. MS Coco). 

- The mathematics in the paper are clear, but I do not like full written words in formulas (like "Softmax", "Smooth", "Conv", etc.). Please replace these by a symbol, which is declared in the text.

Author Response

• If the proposed technique is effective for remote sensing object detection, like demonstrated, it would be interesting to see how well it performs on more general object detection datasets (such as e.g. MS Coco).

Response: Thank you for the suggestion, the proposed method was applied to the Coco dataset and result was included in section 4.4.5 of the revised paper.

 

The mathematics in the paper are clear, but I do not like full written words in formulas (like "Softmax", "Smooth", "Conv", etc.). Please replace these by a symbol, which is declared in the text.

Response: Thank you for your suggestion, we have modified the formula in the article according to your comments.

Reviewer 2 Report

This paper proposes Context-aware Dense Feature Distillation to learn lightweight and advanced detectors for onboard remote sensing object detection. This paper has verified the effectiveness of the proposed method on two datasets. However, my primary concerns are as follows.

 

1. The authors are suggested to show the detection results of different methods visually.

 

2. The authors should compare different methods on more datasets.

 

 

3. The authors should analyze the disadvantage of the proposed method.

Author Response

• The authors are suggested to show the detection results of different methods visually.

Response: Thank you for the advice, the detection results were included in section 4.5 of the revised manuscript.

• The authors should compare different methods on more datasets.

Response: Thanks to your suggestion, we have added the experiments on the COCO dataset and added comparison results with the other two methods, included in the revised version of section 4.4.5. The results show that the method still outperforms all the comparative methods on the COCO dataset, demonstrating the method's validity again.

• The authors should analyze the disadvantage of the proposed method.

Response: There are limitations to the application of CDFD, i.e. it is not suitable for detectors without a feature pyramid network. Therefore, in further research, it is necessary to investigate distillation methods that are more general and suitable for all detectors. The analysis of the disadvantages is included in the conclusions section of the revised version.

Reviewer 3 Report

This mansucript suggests a lightweight model with good performance For remote detection on Cubesats (small satellites).

Previous work shows good performance in remote sensing, but it is difficult to deploy on Cubesats under limited resources. Therefore, this paper proposed a Context-aware Dense Fears Distillation (CDFD).

The proposed CDFD achieves good performance in remote sensing by training a small model from a large model without additional power consumption. In addition, to solve the difficulty of detecting small objects in remote sensing, CFGM was created, and the student model not only mimics the teacher model's feature map, but also extracts rich contextual features to help remote sensing. By integrating the strengths of multi teachers, it shows excellent detection performance with adaptive weighted loss.

I think it's a good paper for light weight application areas.

Author Response

I note that you have made no comments and thank you very much for your acknowledgment of our paper.