Deep-Learning-Based Action and Trajectory Analysis for Museum Security Videos
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThis paper presents a deep learning method for actions recognition in a museum environment. However, there are some shortcomings in the current version:
1. Figure 2 shows the class samples distribution of the action type (normal or anomalous action), grouped by the experimental stage of employment (source or artificial dataset). Some words are imcomplete. This is also true for figure 11.
2. In the experimental section, the Video anomaly detection dataset should be introduced.
3. Real-time video processing system is a contribution. The detailed computation time of each part should be analyzed.
4. Some missing key references about deep learning based video analysis works, such as [1,2]
[1] See more, know more: Unsupervised video object segmentation with co-attention siamese networks, cvpr
[2]Segmenting objects from relational visual data, ieee tpami
Comments on the Quality of English Language
No
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThe research work about action recognition and trajectory extraction is very interesting and valuable. Overall, paper is well organized. I raised few concerns to improve the quality.
1. Do you consider the topic original or relevant in the field? Does it address a specific gap in the field?
2. In “Introduction” section, how does the contribution address a specific gap in the field? Also, the references are missing in the section.
3. The article lacks vital discussion associated to the compared technologies. What is the practical implication to other scenarios or museums?
4. Revise all the abbreviations in the paper and provide full forms for the first time only. E.g. LSTM, YOLO
5. Are the conclusions consistent with the evidence and arguments presented and do they address the main question posed?
6. Tables 2 and 3 are not in sequence. The bold and underline values are multiple, it is better to write the plural form in tables' captions.
7. The development and usage of artificial dataset is unclear.
Comments on the Quality of English LanguageThe English language of the paper is understandable, however, there are areas where specificity could be improved to enhance the overall quality and impact of the paper. There are some grammatical errors in the paper e.g. line 182; so, the paper should be revised carefully. Avoid use of pronoun “they” many times.
The consistency in the writing is poor. It is suggested to revise the complete paper accordingly. E.g. paragraph at line 199.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsThis work presents a cutting-edge video analytics framework dedicated to enhancing security in museum environments. It focuses on two innovative approaches: action recognition for identifying potential threats at the individual level, and trajectory extraction for monitoring museum visitor movements, serving dual purposes of security and visitor flow analysis. Some problems are required to addressed:
Major:
1. Please explain how the re-identification of persons is achieved, especially (as in Figure 9) when there is insufficient ambient light. How is the “Tracking and re-identification of detected people to maintain consistent identities across frames;” (in line 481).
2. How to calculate the relative coordinates between different cameras and their viewpoints in order to achieve re-identification:
3. Please explain the homographic transformations in “section 2.2.1”, and discuss the detail of projecting Figure 5(a) to Figure 5(b).
Minor:
1. It is recommended to use a 'Learning Rate Scheduler' to adjust the learning rate adaptively, which can improve accuracy in most cases.
2. A 360° panorama pictures is recommended to provide.
3. More related work is recommended to cited:
3.1. https://doi.org/10.23919/ICACT56868.2023.10079540
3.2. http://dx.doi.org/10.1145/3406971.3406981
3.3. https://doi.org/10.1109/ICPR56361.2022.9956488
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 4 Report
Comments and Suggestions for AuthorsThis paper proposes a method for video analysis framework designed for security and visitor flow monitoring in museum environments which utilizes existing deep learning methods for action recognition and trajectory extraction.
1) The paper has some issues in its goals. The paper applies existing methods to a domain specific dataset, museum security videos. However, the presented dataset can not benefit future research unless following issues addressed. The dataset is extramely limited in total amount of samples for some classes. More statistical details of the dataset needs to be presented for review. For example, the proportion of multiple individuals in one clip over all and how many rooms have or cameras are used. The uniqueness of the environment of museum needs to be discussed and analyzed. One or two examples of each classes should be given.
2) Two methods are proposed in this work. However, the evaluation of the proposed methods need more results. Authors claim that the investigation successfully identified an validated models that are valuable in a musume setting. However, experiment involving only two models is not convincing to draw this conclusion. More related methods should be evaluated pn this new dataset.
4) The spliting of training and test data is unclear. The test phase shoud unseen data for model evaluation, such as data in different day or with different individuals.
5) The Introduction section and existing models in Sec. 2 demands more references.
6) The meaning of aritifical dataset in figure 1 and 2 is unclear.
7) The limitation of the proposed approach is not discussed. The determination of priority of actions deserve more careful consideration. Well-prepared adversaries may exploit this predetermined priorities to launch an unexpected attack resulting in lost or demange of artworks.
8) It would aid in understanding the proposed method, a general figure with the modules and networks architectures.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 5 Report
Comments and Suggestions for AuthorsIn the paper titled "Deep Learning-Based Action and Trajectory Analysis for Museum Security Videos" introduces an advanced video analysis system for museum security, using deep learning to recognize suspicious behavior and track visitor movements accurately in real-time.
The abstract effectively outlines the problem, proposed solution, and experimental findings. However, it will be good to have some contribution points or (research questions) in the introduction to highlight major contribution.
The methodology is thorough and well-explained, contributing significantly to the manuscript. Overall, with minor revisions, the paper shows promise and can be accepted.
Minor revisions include:
rearranging table captions or number
Images re-ordering
seems like fig 7a and 7b are similar. Probably authors can explain minor difference between them
explain FHD (Line 683)
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsI have no further questions.
Reviewer 2 Report
Comments and Suggestions for AuthorsThe authors have revised the paper according to the suggestions. I recommend to accept in its current form.
Reviewer 3 Report
Comments and Suggestions for AuthorsAll problems have been addressed.
Reviewer 4 Report
Comments and Suggestions for AuthorsThis revision looks good to me.
