Target Detection Method of UAV Aerial Imagery Based on Improved YOLOv5

Round 1

Reviewer 1 Report

YOLO UAV is the proposed approach for target recognition of UAV aerial images using a modified YOLOv5l model. The proposed model improves feature extraction to obtain better performance than the traditional method. 

1. Any specific reason to use only modified Yolov5 large in the proposed method? Because the model's complexity is high in comparison to yolov5 large. Instead of the modified large model, use the smaller and medium yolov5 models.

2. The proposed dataset, didn’t mention any fixed altitude to collect the data. Please mention it because it will have an impact on real-time detection.

3. Could please check out the description of the figures that are not well assigned? 

4. Please check the tables and figures properly.

5. Please check the grammar error.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

This paperdeals with improving detection based on YOLOv5 of small targets on UAV images. Compared with YOLOv5l, the backbone of the proposed method showed Top-1 accuracy of 25 the classification task by 7.20% on the CIFAR-10 dataset. Paper is well-written with a good structure and clear explanation of proposed improvements. Results and comparison with other YOLO versions showed significant improvements while still having quite fast procedure. Language is good and paper is easy to read.

My only suggestions would be following:

- 5 main contributions could be condensed, contributions are in fact various optimisations and upgrades of the YOLOv5 backbone.

- introduction might be shortened a bit (paper is overall quite long).

- also, and more important, I would like to see not only mAP comparison in results section. Authors themselves noted other standard measure indicators and it would be good to see F1, precision an Recall comparison.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

The paper proposes some improvements in the Yolov5 mechanism. The paper is relevant and within the scope of the Remote Sensing journal. However, the paper requires substantial English and grammar editing, and I have the following concerns:

Table 4 is unnecessary. The information in it is redundant.

The authors talk about precision, recall, and f-score, but only use mAP in the paper.

Table 8 is too crowded. I think a better solution would be to change the column names for 1, 2, 3, etc., and explain them in the table legend.

Try putting figure 10 on the same page. The authors could change the position of the (a), (b), (c)... to decrease blank spaces. I recommend changing the style for all figures that are set like this. The blank spaces in between rows are not good, and it would be better to reduce them.

The figures showing bounding boxes should be bigger. It is very hard to see the predictions. Besides, it would be important to show the ground truth bounding boxes. Maybe the authors should include some zoom areas on the predictions for some images.

The discrepancy in the results from Faster-RCNN and Yolov4-tiny is very large and not expected. I suggest the authors redo those models, and even try other backbones in the Faster-RCNN, like the ResNet-101 or ResNeXt-101. Maybe the parameters are not correctly set. Even in the COCO dataset, the differences between the models are not substantially large as shown in here.

The properties and objects of top-view images are very different from ground-view images. It is not convincing why the authors are using the K-means on ground-view images to apply to top-view images.

There are state-of-the-art models that were not shown in this study, such as the Efficient-Det. The authors could compare with more models since nearly all comparisons were variants of Yolo.

 

The paper presents no discussion. The discussion section is very important to highlight how this paper advances in this field and compares it to other studies on similar topics. I recommend the authors make a discussion section.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

Please see the attached pdf document.

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The authors have addressed all my concerns, and this paper looks good now.

Reviewer 3 Report

The authors provided the necessary changes.