An Adaptive Semantic Segmentation Network for Adversarial Learning Domain Based on Low-Light Enhancement and Decoupled Generation

Round 1

Reviewer 1 Report

The manuscript presents an innovative approach to nighttime semantic segmentation, introducing the DLA-Net with notable contributions to low-light image enhancement and segmentation accuracy. However, there are areas for improvement throughout the manuscript.

-- The discussion on the challenges posed by nighttime images and the specific advantages of the proposed DLA-Net could be expanded for clearer context. Additionally, the triplely-adversarial learning strategy would benefit from a more detailed explanation or visual aid to assist the reader.

-- Additional information on dataset characteristics and training parameters would make this work more easily reproducible. Addressing the method's limitations and potential failure scenarios would offer a balanced perspective on its applicability.

-- There are grammatical and syntactical errors throughout the manuscript that make the reading harder. The clarity and consistency of figures and tables could also be improved.

-- A discussion on the broader applications of this research and future directions could underscore the method's relevance and potential for further innovation.

-- The paper would benefit from a more thorough discussion of the limitations of the proposed approach and the potential for scalability or adaptation to other challenging conditions beyond low-light environments

There are several instances of grammatical inaccuracies, awkward phrasing, and inconsistent use of tense.

Author Response

Thank you for pointing these out. Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

The authors introduced a domain adaptive approach for nighttime semantic segmentation that overcomes the reliance on low-light image annotations to transfer the source domain model to the target domain. A low-light image enhancement sub-network combining lightweight deep learning with mapping curve iteration is adopted to enhance nighttime foreground contrast. The body generation and edge preservation branches are implemented generating representations within the same semantic region. In training, a discriminator is implemented to distinguish features between the source and target domains, thereby guiding the segmentation network for adversarial transfer learning. The framework´s effectiveness has been verified testing it on three datasets via metrics: mIoU, PSNR, and SSIM.

 Comments:

 1)    In eq. 1 used authors from ref. 13, please provide the recommendations of selection of parameter alfa that adjusts the size and exposure level of Ie.

  2)     Please correct numerous errors mentioning the numbers before all equations 1-10.

 3)     Please explain for better understanding of potential reader how you select parameter lambda1 in eq. 11 for final decoupling loss. The same comments for parameter lambda 2, 3, 4. You only wrote “In the segmentation network, the hyperparameters l1, l2, l3 and l4 were set to 0.4, 0.5, 20, and 10, respectively”.  This explication is not sufficient to understand the reason for such choose.

  4)     The authors never explained the used performance criteria: mIoU, PSNR, SSIM. It is difficult for potential reader understand the presented in tables 1-5 numerical results without minimal definition of these measures.

Author Response

Thank you for pointing this out. Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

Principal drawback of the revised version that the author did not provide the text highlighting for all corrections in the manuscript that is common practice in the revision process. It is very difficult to understand the authors´ corrections.

This reviewer did not satisfy by the respond on comment 1, There is not any recommendation that explain the selection for parameter alfa value.

This reviewer did not satisfy by the explication (comment 3) that the authors have presented (line 271): “lambda1 and lmbda2 are set to 0.5 and 20, respectively, to balance the scale of losses.” There is not any justification of these chosen values.

Other explication (comment 3) that this reviewer found is as the authors wrote “Since the edge portion is not a large part of the overall image, lambda3 is used to balance the weight of Ledge in Lde, which is set to 0.4 in the experiments” (lines 342). The sentence presented in novel version (page 10, line 358 did not answer on comment 3, this part is identical to the text of previous version. As this reviewer wrote in previous revision “This explication is not sufficient to understand the reason for such choose”.

Author Response

We apologize for the lack of highlighting of changes in the manuscript. We have added highlighting to the changes in this manuscript. Thank you for your understanding. Please see the attachment.

Author Response File: Author Response.docx

Round 3

Reviewer 2 Report

The authors have attended all comments of this reviewer.