MSSDet: Multi-Scale Ship-Detection Framework in Optical Remote-Sensing Images and New Benchmark

Round 1

Reviewer 1 Report

This paper reconstructs a multi-scale ship target detection dataset named HRSC2016-MS based on HRSC2016, which establishes a multi-scale benchmark for ship detection in optical remote sensing images. On this basis, the author proposes a multi-scale ship detection framework (MSSDet) for optical remote sensing images. By designing a joint recursive feature pyramid (JRFP) to extract the semantically strong and spatially refined multi-scale features, the proposed framework effectively solves the problem that the general target detector is difficult to deal with multi-scale targets. Detailed ablation experiments on the proposed HRSC2016-MS dataset and extensive comparative experiments on HRSC2016-MS, HRSC2016 and DIOR datasets demonstrate the effectiveness of the proposed method. Detailed comments are listed in the following sections, which may help to improve this article:

1.In Section 2, because ship targets have large aspect ratio and direction information, using HBBs to detect ship targets may contain unnecessary background information. However, the authors only mention the former (line 127-128) while ignores the latter.

2.In Section 3 algorithm 1, the author uses the symbols d and m in line 8, but does not give the meaning of the symbols.

3.In Section 5, the author mentions that the definition of the evaluation metric mAP is the same as that in the PASCAL VOC 2012 object detection challenge, but does not specify the iou threshold of mAP in the experiments.

4.In the model generalization evaluation experiment, the DIOR scale of the data set used is much larger than the other two data sets, and the author only trains 30 epochs for all models. In addition, single stage models (such as YOLOv3) usually need to train more Epochs to converge. Under this condition, whether the models can converge. It is suggested that the author further explain the above problems.

5.In the ablation experiment on JRFP effectiveness, the author only conducts an overall evaluation on the multi-scale ship dataset HRSC2016-MS, which is not sufficient to illustrate the multi-scale detection capability of the proposed method. It is suggested that the author evaluate small, medium and large targets separately and report the evaluation results, so as to more fully illustrate the multi-scale detection capability of the proposed method.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Please find my comments in the attached file.

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.pdf