Attentive SOLO for Sonar Target Segmentation

Round 1

Reviewer 1 Report

This paper introduces a neural network model, i.e., attentive SOLO, for sonar target segmentation. A new dataset is built to support the task. The experimental results look promising. However, several issues should be addressed:

The manuscript is not sufficiently polished. There are many writing errors, for instance, "segmen- tation", the sub-titles of Sec. 2.3 and Sec. 2.4 are same. I will suggest the authors to proofread the whole paper for several times, and extensively improve the writing.

The related work section is not through, since many relevant instance segmentation works have been missed like Differentiable multi-granularity human representation learning for instance-aware human semantic parsing and Target-aware object discovery and association for unsupervised video multi-object segmentation. 

Many details about the proposed dataset are not provided. It is essential to clarify how the data are collected and annotated, and some statistics of the dataset should also be provided. 

It is surprised to see from Table 2 that the two-stage methods like maskrcnn did not outperform single-stage methods. Please provide some possible reasons to the results of this experiment.

Please provide some failure cases to gain more insights into the method.

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

The authors of this paper used gated fusion-pyramid segmentation attention (GF-PSA) module and reduced noise interference using more semantically strong feature maps for sonar taraget segmentation.

Also, a new sona dataset named STSD is constructed in this paper.

The equations and images are useful to understand Attentive SOLO.

In Experences and Results section, the detail hardware environment and hyperparameters are given, which makes readers can reproduce this work.

 

The authors of this paper used custom dataset named STSD and showed the efficacy of Attentive SOLO with STSD.

If the readers can access STSD, reproduction of this work can be possible.

Attentive SOLO is tested with STSD only.

To emphasize the efficacy of Attentive SOLO, the proposed model should be tested with previous datasets or it is possible to express in figure and tables that STSD is better than previous datasets and sonar target segmentation with STSD is more difficult.

To calculate mean average precision (mAP), several IoU threshold used, but the constraints (such as IoU threshold, detections per image, object size) for computing average recall (AR) are not presented.

There is a minor typo that needs to be changed: the title of section 2.4 should be changed to "sonar datasets".

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

The manuscript approaches an interesting subject, but is too descriptive. Almost one half of the manuscript is referring to the introduction, state of the art, conclusions and references.

I recommend you to extend the manuscript with much more details referring to the applied procedures (which are currently only descriptively presented) and referring to the presentation of the results.

For example, you should detail the applied learning algorithm.  

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The revision has addressed my concerns.

Author Response

Thank you for your comments.

Reviewer 3 Report

OK.

Author Response

Thank you for your comments.