Single-Stage Rotation-Decoupled Detector for Oriented Object

Round 1

Reviewer 1 Report

 

Reviews of the Manuscript ID:  remotesensing-934858

“Single-Stage Rotation-Decoupled Detector for Oriented Object”

 

 

The paper introduces a new single-stage detection approach called RDD (Rotation Decoupled Detector) through adapting an anchor-based method that employs a novel matching strategy. Authors claim that the proposed anchor matching strategy optimizes the way the model learns the objects ‘position without the use of additional rotating anchors. Extensive experiments have been done over three different benchmarks prove that the proposed method reached higher detection accuracy and speed.

Overall, the paper is well organized and written. The contribution is quite interesting and we thank the authors for sharing the link of their code to the community so that it might be re-used for other related tasks. However, more efforts are required to make the idea more clear and the paper more consistent.

Major issues

• Authors applied the proposed oriented object detection to three different remote sensing datasets. We do not understand the utility of mentioning the task of scene text detection several times in the paper while you did not use this application in you experiments.
• The choice of single-stage detector instead of two-stage detector is not motivated nor argued in the text
• The proposed configuration sensors is well described but not argued in the paper.
• Section 3.1 lacks of details. Please provide more details regarding the network architecture. Avoid citing the backbone name in the first time description of your approach. That will make it more general.
• Could your architecture be extended? Or evolutive?
• For example, what is the utility of a dash camera in such dataset?
• Equation (2): Why did not you use N instead of N_pos? Should be an error!
• The notation and the description of equations (3) and (4) are completely missing.

Minor issues

• The paper suffers from minor English issues and authors are invited to improve their scientific writing:
• L31: natural image public dataset --> real life dataset
• L31: delete “and so on” and add and between the previous datasets names.
• L32: write in full letter the first time the acronym in the paper appears like HBB here, OBB (L36), IoU (L40) YOLO(L88), SSD (L90), etc.
• L43: Too early to discuss experiments in this stage.
• L47: a series --> series
• In related-work section, there is no consistence in tenses (past/present)
• L146 and L295: Latex errors
• L158: number of what?
• L191: doesn’tà does not
• L210: is fixed??
• L233: a few --> few
• L250: verification--> validation

 

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

Dear Authors,

This is excellent work! My only concerns are: where is the limitation of this study and the possible future direction!

There is a small error: line 146-147 (possible citation missing).

Good luck!

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

The manuscript is proposing a novel idea and has been written well. The idea is well in line with the scope of the Journal of Remote Sensing.

 

The tone of the manuscript is quite sophisticated that may compromise the truly interested scholars in the field to find it tangible though it professionally describes the developed approach and corresponding technical aspects.

I am wondering about the structure of the proposed bounding box creation in comparison with the available state of the art approaches such as Faster-RCNN. The literature review might be improved in the Introductory Section to further explore how the proposed method may put the science a step forward. Additionally, contrary to the claims raised in several manuscript sections regarding the improvement of the results in the experiments, the empirical pieces of evidence are not well presented.

As such the structure of the proposed method could be demonstrated in further details in addition to Figure 1.

Ultimately, the authors may further clarify on the sentence "The new strategy optimizes the way the model 313 learns the object position information without adding additional rotating anchors" in the conclusion. 

 

Overall, in my opinion, this manuscript's presentation and contents are suitable for publication.

 

 

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

We thank the authors for taking the time to revise their paper through considering our comments. The paper is more organized and ideas are more clarified than the first reviewed version submitted.