Weakly Supervised Learning for Object Localization Based on an Attention Mechanism

Round 1

Reviewer 1 Report

1. The related work should contains researches about weakly supervised object localization, but the author only introduced class activation map and attention mechanism.
2. In the proposed method, the difference between your method and the method in ABN[12] is not clarified. The author should add more illustrations.
3. The whole process of weakly supervised object localization framework is not depicted in this paper, which is confusing why the proposed method can achieve both classification and localization tasks.
4. In the experiments, the feature extractor is too old and some more representative backbones such as ResNet-50 should be compared. Furthermore, the experiments are not thorough.

Author Response

Dear Reviewer,

Thank you for taking the time to review our work.

The answer to your suggestion is attached as a word file.

Best regards,

Author Response File: Author Response.docx

Reviewer 2 Report

This paper presents a very important issue: Explainable Artificial Intelligence, but at this stage this work should be improved significantly.

1. In the section “introduction” please express the area of Explainable Artificial Intelligence (XAI)  - usually understood as the topic concerned with developing approaches to explain artificial systems understandable to human stakeholders. How is it understood in this article?
2. p.28 - you stated that: “In XAI, explainability needs to consider two things: interpretability and complete.” And what about Transparency?
3. A better review of the literature, and more complete technical descriptions should be added and improved.
4. Research results: there is not much explanation about validation.
5. Discussion: This section is poorly structured and does not contain the most important information. There is no comparison to the literature. The authors should explain the difference with other methods found in the literature, it could be important in order to evaluate the performance of the current method to compare it with the state-of-the-art methods.

Author Response

Dear Reviewer,

Thank you for taking the time to review our work.

The answer to your suggestion is attached as a word file.

Best regards,

Author Response File: Author Response.docx

Reviewer 3 Report

The work requires the following corrections:
- first of all the authors should present in the paper the implementation details of the proposed method. All settings, used tools.
- Secondly, tests of the presented model should be performed on several sets of images,
- Next, the process of learning models should be presented in detail, showing the different epochs of learning, indicating whether the process of overlearning does not occur,
- presented results should be average results of several tests on random subsets of data. With presentation of the standard deviation of the results. There can be no conclusion that one method is better than another without statistical verification of such a thesis. Hence, the results presented in the paper are, without thorough verification, at best comparable to other best techniques.
Finally, the authors should make the codes of their experiments available to provide a reference point for other authors.

Author Response

Dear Reviewer,

Thank you for taking the time to review our work.

The answer to your suggestion is attached as a word file.

Best regards,

Author Response File: Author Response.docx

Reviewer 4 Report

This article presents a method for adjusting the weights of the classifier
to solve the suboptimal problem more efficiently than the existing method.

The article presents some weaknesses that the authors should address to improve their valuable results.

1. The introduction provides a brief overview of the study.
However, as the problem is well framed,
it is not clear what unique challenges are associated with the task.
The introduction should contain more details about the research problems.
In addition, the introduction should clarify the contributions
of this paper on how to address the research challenges and the contributions of the authors.
I suggest that the authors emphasise this point more.

2. Section 2, related works, lists the literature employing CAM and AM techniques from different authors.
However, the related works section does not indicate the research gap between this work and the other works.
Has this work addressed any limitations of existing computer vision techniques?
Please add some details on this.

3. Section 3 is presented very well and the method seems sound.
However, I think the authors should highlight two main aspects:
  (a) the research gap between this work and the limitations of other existing work;
  (b) the motivations that led the authors to choose the methods described.

4. In section 4:
  a) why did the authors employ VGG-16 or Inception-v3 as a feature extractor? Please give reasons.
  b) A comparison is presented in Table 1. What is new in VGG-Ours and Inception-v3-Ours?

5. In section 4, I suggest adding a brief discussion on timing.
In particular, as the authors indicated "There is no pre-processing task,
such as hiding part of an image, required before training the model.
So it does not take extra time to train the model." as one of their contributions might be useful to know:
  a) how long does it take to perform the training?
  b) How long does it take to perform the classification?
  c) In table 1 it might also be useful to add the times of the comparison methods,
  if available.
Solving this point could be significant in such a task. Furthermore, the answers
to these questions would be valuable both for the manuscript itself and for the
for the community to better understand the performance.

6. I suggest, in Section 5, adding a brief discussion in which the authors should
provide a summary of their main findings and highlight the advantages of their method.

Minor issues:
- The manuscript is quite short and readable, my compliments.
However, there are some typos, or different font types along the manuscript.
Please revise it accordingly.

Author Response

Dear Reviewer,

 Thank you for taking the time to review our work.

The answer to your suggestion is attached as a word file.

Best regards,

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

The changes have been accepted.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

The authors have significantly improved the text given in the manuscript. 

Reviewer 3 Report

Thank you for making the amendments. In my opinion, they are not sufficient for the final acceptance of this work. It should be added to it:
- visualization of the learning process, individual epochs, with showing the effectiveness of the classification and loss function,
- the codes used in the experimental part of the paper should be added to create a reference point for other authors,
- the introduction lacks reference to the historical background leading to the development of current techniques,
- in order to save the reader's time, it is necessary to describe in great detail the data used, give information on how they are plugged into the model, how to download them, from where, and what results on these data other authors have had.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 4 Report

The authors have addressed all my suggestions, therefore I suggest to accept it in its current form.

Best regards

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 3

Reviewer 3 Report

After the second round I believe my comments have been addressed and I have no further remarks to make.