An Efficient Detection Framework for Aerial Imagery Based on Uniform Slicing Window

Round 1

Reviewer 1 Report

In this paper, aiming at the common problems of UAV detection, an improved tiling reasoning framework of uniform sliding window is proposed. specifically, the mixed data strategy is used to develop reasoning and training pipeline. scale filtering is used in global and local detection of original images and patches to keep scale invariance. Although this method has been very effective for real-time detection of UAV images, there are still several main problems in the manuscript:

1. Section 3.1 mentions that the image is cropped with the same length-width ratio as the original image during training, in order to use a single model for local and global detection.But section 3.2 proposes random scale jitter during random clipping, so how will they be trained in parallel?If zooming a batch of images with the same scale will result in different scaling factors for each image, how will the Scale Filtering proposed in Section 3.3 be implemented?

2. The comparison method in Table 1 uses different detectors and backbone, and the comparison seems unconvincing.Almost all the methods proposed in this paper are data preprocessing methods, whether we can use the data preprocessing methods in other methods to prove the effectiveness and versatility of this method.

3. In Table 3, we see the results of some experiments.However, the benchmark method in which experiments 4 to 10 were extended is not given.In addition, the author did not conduct sufficient ablation experiments between these methods, and did not analyze the interaction between them.

4. In figure 5, the author only gives the final test results, which is not convincing. The detection results of the benchmark method should be given, and then the detection results of the improved method should be given. finally, the test results should be compared to analyze the effectiveness of the proposed method.

5. In Table 1, some methods are compared, but they are a bit out of date and need to be compared with the latest methods.

6. In section 5.5, some other real-time detection methods should be compared.

 

 

No comments

Author Response

Dear reviewers,

Thanks so much for reviewing our manuscript! We are writing to express our sincere gratitude for the time and effort you have devoted to our manuscript "Efficient Detection Framework for Aerial Imagery Based on Uniform Slicing Window" for Remote Sensing. We appreciate for your pertinent comments and valuable suggestions, which help us clarify and strengthen our work.

We have carefully considered all of your comments and have made the necessary revisions to address them. We believe that these changes have significantly improved the manuscript. The notes of major revision and detailed replies referring to your concerns and questions are described later.

Once again, thank you for your contribution to the peer-review process and for helping us to improve our work. We appreciate your expertise and dedication.

 

Kindly regards.

Yong Song*

Author Response File: Author Response.pdf

Reviewer 2 Report

This paper presented a method for object detection in UAV images. The idea sounds clear. In general, the paper is well-structured and -written. Authors gave promising results on both benchmark datasets and real-world datasets. I only have some minor concerns:

1) for object detection in UAV images, many methods have been examined in literature, authors may oversee this: Zhang, H., Sun, M., Li, Q., Liu, L., Liu, M., & Ji, Y. (2021). An empirical study of multi-scale object detection in high resolution UAV images. Neurocomputing, 421, 173-182. Authors may further discuss the effects of high-resolution images on the results.

2) in real-world testing, the experimental setup is unclear. Authors should provide the height of the UAV and other settings. 

3) I also suggest authors further discussing the categories of different objects. Actually, it is related to the generalization of your method on different categories. 

4) At last, authors may add the model size of your method in comparison to other methods. 

 

Author Response

Dear reviewers,

Thanks so much for reviewing our manuscript! We are writing to express our sincere gratitude for the time and effort you have devoted to our manuscript "Efficient Detection Framework for Aerial Imagery Based on Uniform Slicing Window" for Remote Sensing. We appreciate for your pertinent comments and valuable suggestions, which help us clarify and strengthen our work.

We have carefully considered all of your comments and have made the necessary revisions to address them. We believe that these changes have significantly improved the manuscript. The notes of major revision and detailed replies referring to your concerns and questions are described later.

Once again, thank you for your contribution to the peer-review process and for helping us to improve our work. We appreciate your expertise and dedication.

 

Kindly regards.

Yong Song*

Author Response File: Author Response.pdf

Reviewer 3 Report

This paper provides an Efficient Detection Framework for Aerial Imagery Based on Uniform Slicing Window. This is an interesting and technical research paper that makes a valuable contribution to the Detection Framework for Aerial Imagery framework. The authors have conducted substantial research to draw meaningful conclusions from their findings. Furthermore, I find the practical applications of their research to industry to be particularly noteworthy. Regarding this idea, I have a few questions:

1. Discuss the YoloV8 model's success in object detection in a related work section.

2. Better to write the method name instead of numbering in the contribution list (3rd point).

3. What is the difference between Mosaic Plus Augmentation and Normal Mosaic  Augmentation?

4.  Why did you design two extra combination manners in Mosaic?

5. What is the difference between Crystallization Copy-paste Augmentation and Mosaic Plus Augmentation?

6. remove the hashtag from the number in section (5.4.1)

7. Better to add state-of-the-art papers and include these relevant papers in the list.

I. https://shorturl.at/fHK68

II. https://shorturl.at/djqV7

III. https://shorturl.at/gNQ16

Please improve the reading flow. Some paragraphs need a connection with the previous paragraph. Thanks 

Author Response

Dear reviewers,

Thanks so much for reviewing our manuscript! We are writing to express our sincere gratitude for the time and effort you have devoted to our manuscript "Efficient Detection Framework for Aerial Imagery Based on Uniform Slicing Window" for Remote Sensing. We appreciate for your pertinent comments and valuable suggestions, which help us clarify and strengthen our work.

We have carefully considered all of your comments and have made the necessary revisions to address them. We believe that these changes have significantly improved the manuscript. The notes of major revision and detailed replies referring to your concerns and questions are described later.

Once again, thank you for your contribution to the peer-review process and for helping us to improve our work. We appreciate your expertise and dedication.

 

Kindly regards.

Yong Song*

Author Response File: Author Response.pdf

Reviewer 4 Report

The paper focuses on drone object detection and the obstacles faced in this domain. The proposed approach involves using a mixed data strategy for inference and training. Global detection is performed on the original image to handle large objects and prevent truncation, while local detection is simultaneously conducted on corresponding sub-patches to enhance tiny objects. The training data includes both original images and patches generated by random online anchor-cropping to increase diversity. To address objects of different scales, the authors use scale filtering to assign them to either global or local detection tasks, ensuring scale invariance and optimal fused predictions. The tiling inference is executed in parallel, maintaining high efficiency suitable for real-time applications. The study introduces two custom augmentations specifically for tiling detection, creating more challenging drone scenarios and generating a greater number of valid annotations, especially for rare categories and overlapping objects. Through experiments on public benchmarks and real-world drone images, the proposed tiling framework demonstrates its effectiveness in improving general object detection in drone scenarios. Moreover, the code is available which makes reproducing the results possible.

I have the following comments to be addressed:

1. In Table 1, is it fair to compare methods with Resnet50 backbone with methods that use CSP-DarkNet53 as

 

3. In Table 4, could report the inference and process time for the first three to make the table complete?

 

4. On your tables, please also report the standard deviation values to make the comparisons statistically meaningful. 

 

 

 

 

Author Response

Dear reviewers,

Thanks so much for reviewing our manuscript! We are writing to express our sincere gratitude for the time and effort you have devoted to our manuscript "Efficient Detection Framework for Aerial Imagery Based on Uniform Slicing Window" for Remote Sensing. We appreciate for your pertinent comments and valuable suggestions, which help us clarify and strengthen our work.

We have carefully considered all of your comments and have made the necessary revisions to address them. We believe that these changes have significantly improved the manuscript. The notes of major revision and detailed replies referring to your concerns and questions are described later.

Once again, thank you for your contribution to the peer-review process and for helping us to improve our work. We appreciate your expertise and dedication.

 

Kindly regards.

Yong Song*

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

there are no further comments and I think it can be accepted.