SIVED: A SAR Image Dataset for Vehicle Detection Based on Rotatable Bounding Box

Round 1

Reviewer 1 Report

In this paper, the authors provide a dataset for SAR vehicle detection, which is a meaningful work. Here are some problems to be solved and then this paper should be reviewed again.

1.    Please give more information about how the raw data is obtained.

2.    Add related work about SAR detection datasets, such as ls-ssdd-v1.0: a deep learning dataset dedicated to small ship detection from large-scale sentinel-1 sar images, sar ship detection dataset (ssdd): official release and comprehensive data analysis, and so on. The authors should discuss different kinds of object detection in SAR images, e.g., ship detection, airplane detection, video SAR shadow detection, e.g., shadow-background-noise 3d spatial decomposition using sparse low-rank gaussian properties for video-sar moving target shadow enhancement.

3.    Please provide more necessary information about the data set, including the source of SAR images, shooting location, satellite name and other information.

4.    How is the urban scene figures cut into a uniform size?

5.    Why did the author put the MSTAR slice into the dataset? There are significant differences between MSTAR slices and urban scene figures.

6.    In this work, why do the authors use semi-automatic annotation instead of manual annotation? Is semi-automatic labeling more reliable than manual labeling? Labeling more than 1000 slices does not seem to be a huge project.

7.    In semi-automatic labeling, what is the basis for selecting parameters of CFAR? Will this have a serious impact on the correctness of dataset construction?

8.    Please analyze the advantages of your dataset.

9.    What is the significance of using YOLOv5 network?

10.  Please add at least three more models in the baselines.

11.  The authors should consider and think about it and add them in this work ending these and it is interesting i.e. a mask attention interaction and scale enhancement network for sar ship instance segmentation, htc+ for sar ship instance segmentation, a polarization fusion network with geometric feature embedding for sar ship classification, balance learning for ship detection from synthetic aperture radar remote sensing imagery, high-speed ship detection in sar images based on a grid convolutional neural network, depthwise separable convolution neural network for high-speed sar ship detection, a full-level context squeeze-and-excitation roi extractor for sar ship instance segmentation, a mask attention interaction and scale enhancement network for sar ship instance segmentation, hyperli-net: a hyper-light deep learning network for high-accurate and high-speed ship detection from synthetic aperture radar imagery and so on.

12.  IMHO, the Conclusion should be re-written to 1) explicitly describe the essential features/advantages of the proposed method that other methods do not have, 2) describe the limitation(s) of proposed method, and 3) what aspect(s) of the proposed method could be further improved, why and how.

13.  The English should be improved greatly.

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 2 Report

The paper presents the construction of a new dataset for SAR image vehicle detection, named SIVED, in Ka, Ku, and X bands of data. The authors propose an algorithm for automatic annotation to improve efficiency and employ rotatable bounding box annotations to improve positioning accuracy. The dataset is rich, stable, and challenging, comprising 1044 chips and 12013 vehicle instances situated in complex backgrounds. The authors evaluate five detection algorithms on SIVED to establish a baseline, and the experimental results demonstrate that all detectors achieve a high mean average precision (mAP) on the test set, highlighting the dataset's stability. Overall, the paper contributes to filling the gap in SAR image vehicle detection datasets and demonstrates good adaptability for the development of deep learning algorithms. The proposed dataset will be valuable for researchers interested in developing SAR-based vehicle detection algorithms.

 

1.     Will the dataset be made public? If so, when and where to download the dataset?

2.     The detection results reported are deep-learning based methods. How about the detection results of CFAR?

3.     Some typos and grammar errors need to be checked and rectified.

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 3 Report

This article construct a SAR image dataset named SIVED, which owes three properties: richness, stability, and challenge. Overall, the structure is complete, and the theory and experiments are convincing. However, there are still some crucial problems that need to be carefully addressed before a possible publication. More specifically,

1. Please clarify how to splice 16 MSTAR images into 4×4, random or according to the order, if the random composition can increase the number of chips of MSTAR and further expand the SIVED dataset.

2. The accuracy of the Semi-Automatic Annotation of SIVED mainly depends on the first step: Algorithm Automatic Annotation and the mask size will directly affect the threshold T. So how is the 30 obtained?

3. Please explain the meaning and function of the positive samples in Line 219.

4. To better prove the superiority of the SIVED dataset, it is best to add a set of experiments to compare the performance of the network trained by the constructed dataset and the original dataset (FARAD、MSTAR), respectively.

Author Response

Please see the attachment

Author Response File: Author Response.docx

Round 2

Reviewer 3 Report

Thank the authors for the modifications and responses made to the paper based on my comments. I personally believe that the paper can be considered for acceptance.