AOGC: Anchor-Free Oriented Object Detection Based on Gaussian Centerness
Round 1
Reviewer 1 Report
The paper is focused on anchor-free oriented object detection method based on the Gaussian centerness(AOGC),
The approach is well explained and the example allows readers to understand the whole problem, and also to replicate it.
Despite the paper is well presented and structured, I suggest introducing the main results in the abstract in brief, so that it is possible to have an idea about the overall ones.
Moreover, the introduction, in the presented form, is quite negligible. I suggest
restructuring it explaining in-depth the maintenance problem.
The English form has to be revised for the presence of some typos. The literature review appears not completely updated.
good
Author Response
Please see the attachment.
Author Response File: 
Author Response.docx
Reviewer 2 Report
This paper proposes anchor-free oriented object detection method based on the Gaussian centerness(AOGC), a single-stage anchor-free detection method. AOGC achieves mAP of 74.30% on the DOTA-1.0 datasets and 89.80% on the HRSC2016 da-tasets, respectively. AOGC exhibits superior performance to other methods in oriented anchor-free object detection methods.I consider that the work is interesting, and the conclusions obtained have certain application potential. Adequate revisions to the following points should be undertaken to justify the recommendation for publication.
1. The abstract of the paper needs further revision and language polish.
2. Many abbreviations in the text need to be given full names.
3. Check the mathematical notation especially for the proposed method.
4. The format of the mathematical formulas in the article is not standardized, please check and correct them one by one.
5. The following papers published in related Journal may be related to this manuscript, we would very grateful if you could cite this paper, e.g., "DTCSMO: An efficient hybrid starling murmuration optimizer for engineering applications", "HG-SMA: hierarchical guided slime mould algorithm for smooth path planning".
6.The discussion of the results needs to include the strengths and weaknesses of the proposed algorithm.
7.There are some grammatical and typographical errors in the paper. Read carefully. Need to improve all.
Moderate editing of English language required.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 3 Report
This paper deals with the problem of oriented object detection. Particularly, an anchor free approach is designed. Three critical components are introduced, including a contextual attention FPN to capture contextual information of the target, a label assignment module to predict the orientations, and a Gaussian kernel based centerness branch to determine the significance of anchors. Experimental results demonstrate the effectiveness of the proposed method. The writing and structure of the manuscript is good. However, I have the following concerns:
1. The overview structure of the proposed method needs to be further clarified. Particularly, how the label assignment module is incorporated with the deep network. Figure 3 should be modified.
2. The performance improvement of the proposed method in comparison with state-of-the-art methods is not significant.
3. The effectiveness of OLS in ablation study should be investigated.
4. The investigation of parameter settings in loss function is missing.
The writing and structure of the manuscript is good. The paper reads well.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 4 Report
This article proposes and evaluates an anchor free object detection method suitable for oriented objects in remote sensing images. Its main contributions include an extension to feature pyramid network (FPN) and support for angle of rotation of the detections.
Strengths:
1. The paper is well written, easily understandable despite some unusual English formulations.
2. The proposed techniques, Contextual Attention FPN (CAFPN) and object bounding box label assignment (OLA) and Gaussian centerness branch (GC), seem to be effective and leading to improved detection performance with respect to standard metrics such mAP.
3. Experimental evaluation is thorough in the sense that the performance of the proposed system is studied on two data sets including ablation studies. In addition the performance of the proposed system is compared with many previously published systems.
Weaknesses:
1. Experiment results do not include statistical significance tests. Some improvements seem marginal (Table 3). It is better to state this fact for the benifit of the readers.
2. It is not clear in which situations the proposed system is better than existing systems (Table 2, Table 3) It may be useful for the readers if there is some criterion specified to pick AOGC over other competitive systems. For example, a table comparing FLOPS of those competitive systems with AOGC would have been useful.
Other minor issues:
1. Line 309: Matrix A represents diagonal matrix of eigen values of which matrix?
2. Define L_cls in equation 9
3. Table 1 is split in two pages
4. Define row headings of Table 2
English quality OK, but can be improved some minor errors/style issues
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Round 2
Reviewer 3 Report
The authors have revised the manuscript according to the comments, and the paper has been improved accordingly.
An additional suggestion:
The OLA module is related with the design of the loss function, and should not be contained in the section of network design.
The sturcture of the manuscript is reasonable, and the paper reads well.
Author Response
Thank you for pointing this out. We agree with this comment. The network design section really shouldn't include our OLA module, so we have changed the title of section 2.2.1 of this article from Network Architecture to Overall Architecture. It makes sense to include an introduction to our OLA module in Section 2.2.1. We also changed the title of Figure 3 accordingly.
