Multi-Level Fusion Model for Person Re-Identification by Attribute Awareness

Round 1

Reviewer 1 Report

Authors proposed a multi-level fusion person re-identification network to fuse global and local features of pedestrians. The network introduces the instance batch normalization module and is transformed by a non-local method based on attention enhancing the performance of feature extraction. With bidirectional training of attribute- and identity-based networks, the network can learn the relationship between attribute labels and pedestrian features, as well between identity labels and pedestrian features. The method was tested on two standard datasets in the state of the art that are Market-1501 and DukeMTMC-reID and its effectiveness was verified by comparison with the latest methods, showing an outperformance.

The article is well written and interesting since the topic is really relevant today where cameras are ubiquitous for several critical applications. 

Time complexity should be better explained, with a focus on the architectures used for training ad with indication about inference time on selected hardware in order to give a practical idea about the feasibility of the proposed approach for example in real-time.

As a minor comment there is a formatting issue on page 18 on the "Data Availability Statement".

For the reasons I highlighted I believe this manuscript deserve a publication, after a proper revision of the issue reported by the comments above. 

Author Response

Thank you for your review. I have revised formatting issue on page 18. And I have revised the manuscript.

Author Response File: Author Response.docx

Reviewer 2 Report

Please find the comments below:

• In the abstract, please mention the novelty of the work.
• Please mention why it is different from other established procedures to publish this article. Also, in the abstract please briefly highlight the author’s findings for better clarity.

• The introduction is well written; however, I will recommend inserting some points to improve the clarity-
  1. Line no 19- why Pedestrians are rigid and flexible characteristics-Justify more!
  2. Did the author think of using an Artificial Intelligence modeling system for this kind of image-based retrieval technology?
  3. Line no 23- “The main idea of traditional image-based person re-recognition is to compare the similarity of two identities”-this can be done by AI-based algorithms then why this technology is needed?
  4. Line no 26- “slightly rough-What does it mean – please be clear with your statement and I will recommend using more scientific words-rephrase it
  5. Line no 41- more attributed do not imply better performance-why! Justify
  6. Line no 64- What the solution for these problems in the future-A future direction must be mentioned here
  7. What is new in this study? -please clarify at the end of the introduction part.
• Increase the clarity of figure 1, some words are not readable. As per journal criteria, figures must have 600 DPI, please make sure all the figure has good pixel quality
• Line 222- what is the link between loss function and person’s re-identification dataset-not clear.
• Explain figure 3 more clearly and justify specifically to make general readers understand well.
• Figure 5 needs to be changed. The small title box (Lin, Yin, and MLAReID) is overlapping inside the main figure-this is not the correct way to represent it- please show the title box outside/beside the graph.
• The same remarks for figure 6 also
• The same remarks for figure 7 also
• Same remarks for figure 8 also; and explain about Epoch and mAP in the figure legends.
• Same remarks for figure 9 also- not readable

• The conclusion part is in very poor shape, needs to be rewritten properly. The authors have mentioned all general information but what is the relation of the findings and please mention what are the challenges and limitations of the mentioned process has. Also, mention some future direction/probable future studies based on these proposed studies

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

I am rather happy with the main technical elements of this submission. The problem addressed is relevant, topical, and challenging; the method proposed is sound, sufficiently novel, and interesting; the evaluation is strong, with sufficient evidence as regards the performance of the proposed method, comprehensive comparisons, etc.

However, the presentation of the material needs a lot of improvement. Even the first couple of sentences show how lax and imprecise the wording is:

• "Image-based person re-identification can be transformed into a similar image-retrieval problem." - What does this mean exactly? How is the similarity of an IR problem with a re-identification one measured? The authors do not in fact mean "similar" but rather "equivalent". Alternative, even better phrasings of the entire sentence are also possible.
• "Few identity-based methods consider pedestrian attributes, and many methods that consider pedestrian attributes and identities fail to fully activate their relationship." Here we again have an example of poor and vague language. How does one activate a relationship? This means nothing. Again, you are after a very different word "exploit"

The manuscript is full of problems like this which make it hard to follow and which break the flow of content. The authors should make sure that significant improvement is effected if I am to recommend this for acceptance. I also suggest that the authors comment on the potential use of meta-learning on top of their method (which is in any case to the authors' benefit as it contextualizes the contribution better) as described in e.g. "Learnt quasi-transitive similarity for retrieval from large collections of faces" (CVPR)

Author Response

Thank you for your review. I have revised the manuscript. Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

I am happy with the authors' latest revision - well done.