Intelligent Video Surveillance Systems for Vehicle Identification Based on Multinet Architecture
Round 1
Reviewer 1 Report
The paper review existing AI-based technologies for video surveillance. Although the paper is interesting, its novelty is not clearly depicted:
Firstly, the provided review is not systematic – the Authors do not provide any information on the process of literature selection, so this process and the results look subjective.
Secondly, the purpose of the provided review is not clear. Frequently, the review is conducted for discovering some gaps, under-researched areas, tendencies and so on. The significant part of the paper is devoted to description of existing solutions/algorithms, and this is difficult to make conclusions on the topic.
Thirdly, the paper includes results of some experiments with existing models and a proposed model (combination of 4 models). Again, there are some interesting results here, but their novelty is not highlighted (or just small and limited by computation time). Empirical properties of the proposed model seem not analysed/well described.
Taking into account the points above, the article is a mix of a popular paper for unaware users, a literature/technology survey, and a research paper. In my opinion, if the Authors will focus the paper to one of these three components (probably, split it and provide more details), it would be much easier for the Reader to follow the corresponding novelty and logic of each part.
Author Response
Please see the attachmen
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The article starts as a survey that initially appears to target a larger window of computer vision applications, but then focuses on license plate identification. Eventually, the study even puts forward an approach composed of 4 different models that is applied on the authors own data set which imitates different realistic environments.
Unfortunately, I do not find the article worth publishing. The focus of the article is not clear. The survey part is rather weak, there is no clear conceptual view of the domain, but more like some ideas compiled from different articles. It has the look of a seminar work. Moreover, the English is poor, there are numerous grammar mistakes in the manuscript and this is a great disadvantage for a survey. There are (too) many Figures with real-world pictures.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Reviewer 3 Report
This manuscript proposed a method for vehicle identification and tracking. However, the proposed method is only a simple combination of existing models for object detection, ReID, and OCR. This manuscript does not propose methods for improving the models or for combining the models. The result is simply a re-measurement of the accuracy of the existing models. So, it seems that this paper does not make a contribution for vehicle identification and tracking.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Reviewer 4 Report
The article intends to look for up-to-date solutions for vehicle identification tasks (combining re-identification with license plate reading) under variable operating conditions. Besides that, the authors provide a vast review of recent works on the topic.
However, the article contains a number of issues to be discussed.
(1) The title should correspond to the content of the paper and contain the key message. It sounds too general in its current form.
(2) The results achieved by the research and theoretical values are expected to be emphasized in the abstract. The abstract can be made more detailed by including metrics from their findings and inferences (drawn from experimental results). It should be more specific. Based on the current version, it is not clear what system is introduced in the work.
(3) The manuscript should be spell-checked and grammar-checked. The quality of English must be improved.
(4) In the conclusion section, the author should provide a clear scientific justification for this work and indicate use-cases and extensions if appropriate. It should sum up the whole research based on the obtained outcomes. Please do not use any square brackets (references) in this part.
(5) Tables 3&5 require more comparison with analogs. Why do these tables contain a single net (system)?
The introduction must be restructured. It looks reasonable to split this section into several ones like Introduction, Related Works, Theoretical Background, and so on.
The introduction should be rewritten/updated for stating i) what the problem is; ii) why the problem is relevant (a motivation); iii) what the work's contributions are, and iv) how the contributions relate to the relevant literature.
(6) The paper's structure must be restructured. The manuscript contains lots of exuberant data. The authors should think about the content's optimization.
(7) The motivation is extremely vague. It must be formulated clearly.
(8) The discussion section is not sufficient based on the provided tables. There must be more analysis on the grounds of the obtained results.
(9) The authors do not explain why they have chosen these specific metrics for this task. It should be explained what these metrics mean and why they were chosen.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
Unfortunately, I do not find the manuscript improved enough to be published, although there are some clear improvements as compared to the previous version. The focus seems still to be split into too many directions.
There is now a "multinet" architecture that appears in the title, but the term does not appear in the body of the article. I think that the article concentrates on too many tasks at the same time, instead of focusing on a single task and making a thorough study with it.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Reviewer 3 Report
The revised manuscript presented the purpose of a proposed method, and the results were also improved. However, there are still many ambiguous and non-detailed descriptions. So It is necessary to make the description more clear and detaily. In addition, you should explain the contribution to the field of vehicle identification by comparing with previous vehicle identification (and re-identification) methods
-
In Abstract, it was argued that the previous methods of vehicle identification are unreliable in certain situations, but the certain situations weren't presented. Please add the explanation of the corresponding situation. In addition, the last sentence “Through this article .. as the starting point for this research” seems to be out of line with the main topic.
-
(the last sentence of the first paragraph in Section 1.4) What does the situation mean from “However, there are many situations where these cameras do not work properly”? Please provide a suitable example for the corresponding situation.
-
In Section 1.4, The third paragraph (“Vehicle tracking for security requires … to other different solutions shown here.) seems to be a main goal to be proposed in this manuscript. However, the content is written after several sub-sections. So the reader may find it difficult to understand the main purpose. It seems that this content should be dealt with in the first section of the manuscript (Introduction Section).
-
In Figure 2, the meaning of this figure is ambiguous. Please modify this figure or add the description of an example of applying DORI to the picture.
-
YOLOv5, an object detection model, is applied to recognize the license plate number. However, State-of-art OCR engine can be also used to recognize the license plate number enough such as Aster (doi:10.1109/TPAMI.2018.2848939) or TextSpotter (doi:10.1109/CVPR.2018.00527). What are the advantages of using YOLOv5 for plate number recognition compared with the OCR engine?
-
In Section 5.1.2, it seems that the term “custom trained YOLOv5” means the YOLOv5 trained by the custom dataset. In this case, the term “custom” may be unsuitable. The term “custom” is usually used to mean that the network structure has been modified.
-
In Section 5.1.3, is the feature map extracted in the license plate detection step used to extract the visual characteristics of vehicles in re-identification? or are the two things separate? If the re-identification uses the feature map from the license plate detection, please describe this clearly. Otherwise, describe in detail the process of extracting the visual characteristics.
-
In Section 5.2, please cite the source of the used dataset if you use a public dataset. or state that you used your own data set.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Reviewer 4 Report
The authors have considered the remarks.
The paper may be accepted now.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Round 3
Reviewer 2 Report
I appreciate the efforts of the authors to improve the manuscript.
There are still some typos, please recheck the manuscript. For instance, in the abstract, line 10, there are two dots ending the first sentence. Similarly, see line 61. Line 366 has a reference to [3737].
Check the formatting for lines 730-738.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
