DSA-SOLO: Double Split Attention SOLO for Side-Scan Sonar Target Segmentation

Round 1

Reviewer 1 Report

line 93-95- why can't semantic segmentation determine the category of the target? are all of them binary segmentation techniques? ideally it should be trivial to make the last layer a classification layer and also determine category of each pixel, so can you please clarify?

please provide slight intuition on what is spatial and channel information with respect to sonar scan images

out of all the different instance segmentation frameworks what was the motivation to use SOLO as a base for the proposed DSA-SOLO?

deeplabV3 is also a one stage instance segmentation technique however a critical review for this technique has not been provided, advise authors to include a line or two highlighting the disadvantage of deeplabV3 also

There seems to be a focus to demonstrate that DSA-SOLO is a better model that SOLOv2, however experimental results have been validated only on sonar data, request authors to also check on benchmark RGB image instance segmentation datasets 

Utilized dataset seems to be significantly smaller in size compared to existing RGB instance segmentation datasets? How did the authors adapt to this low volume data

what is the train/validation/test split? was any finetuning/transfer learning strategies used?

Please include more details on the data processing pipeline, i.e data input size, sequence of transformations before being fed to model and final model output shape and size

Do authors intend to open source the code so that the results can be reproduced?

advice authors to also report GFLOPS (model complexity) if possible, and number of model trainable parameters 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

The topic undertaken by the authors is very interesting and important from many points of view. However, the article contains some errors. Therefore, in order to improve the scientific quality of the article, some corrections should be made, as listed below:

1. Is Figure 1 the authors' work? In addition, every figure should have a title, not a sentence description.

2. In lines 37-38 the authors write about research related to sonar image target segmentation - please provide the appropriate references.

3. I propose to title Section 2 as "Literature Review". It should also be a bit more extensive.

4. Also the section "Conclusions" should be slightly more elaborated - eg there is no indication of research limitations in it.

5. Please diversify your references more - so that they reflect research conducted around the world.

6. I suggest to read the entire manuscript carefully again and correct some stylistic or editorial errors.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf