Monocular Depth Estimation from a Single Infrared Image
Round 1
Reviewer 1 Report
This paper proposes a self-supervised monocular depth estimation method for infrared images. Both quantitative and qualitative experiments validate the proposed algorithm. Some comments are listed as follows:
- In the abstract, the authors mention that they propose the framework. This loss solves the problem of appearance matching loss… The statements are very confused. Please clarify these. And many losses are mentioned one by one. It would be better to summarize them and put the main contributions in the abstract.
- Dose the method to generate Pseudo-Label affect the performance of the proposed framework?
- Table 1 is not mentioned in the full texts.
- Related works about deep learning-based infrared image processing tasks are recommended to be reviewed, including Single infrared image enhancement using a deep convolutional neural network, Infrared pedestrian detection with converted temperature map, etc.
- It is suggested to further improve the presentation. For example, some tables are too large, including Table 1 and Table 2.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Pros:
• The article presents a valuable approach to the self-supervised monocular depth estimation based on thermal images.
• The proposed idea is quite straightforward, but authors claim that it is effective.
Cons:
* Line 5: |This loss solves the problem of appearance..." - what loss?
• Lines 33-36: „A thermal camera records images from an object’s radiation, which is unaffected by changes in the external environment. ” - what about the radiation reflected and from the other sources, i.e. from the sun or other warm objects?
- Lines 67-68 "We show our proposed Self-Guided framework performance qualitatively and quantitatively..." - weak English
• Line 115: 3.1. „Review of self-supervised learning using thermal” thermal what? images, cameras?
◦ The title of this subsection does not match the content. The subsection consist only one reference and presents a standard approach of training process not a review.
◦ In my opinion, it would be better to separate a description of the approach proposed by the authors and the related work (e.g. in two chapters).
- Line 175 - please add an explanation how do you define low- and high-level pixels.
• Table I appears too early, before chapter 4.
• Tables I and II are truncated.
- Line 216: "4.4. Ablation Study" The purpose of the ablation study is not entirely clear.
• Line 220: lack of reference for DIFFNet.
- "GT" abbreviation in Fig. 3 is not defined.
• Table 3, It has not been explained why such CNN architectures were selected for the experiments (i.e. GBNet, DIFFNet).
• greater, smaller →should be changed to higher, lower.
• In my opinion the authors should compare their results from chapter 4 with the results of other solutions presented in the literature, or at least explain why they do not do it.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
The authors have addressed some of my concerns. But there are several points that need to be clarified.
“such as object detection, segmentation, image enhancement, person re-identification, and visual localization” each task need one citation, as my previous comments.
Different methods for generating pseudo label can be tried.
Some figures are still too Large.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 2 Report
I accept the changes made by the autors.
Author Response
Thank you for reviewing our paper.
Round 3
Reviewer 1 Report
No further commments
