Attention-Guided and Topology-Enhanced Shift Graph Convolutional Network for Skeleton-Based Action Recognition

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

In this manuscript, an innovative and novel model named attention-guided and topology-enhanced shift graph convolutional network (AT-Shift-GCN) is proposed, which is a breakthrough in skeleton-based action recognition that provides a lightweight model with enhanced performance on three datasets. The article inclueds all relevant references, and descirbes the AT-Shift-GCN clearly with fine quality english language. 

In order to evaluate the performance of AT-Shift-GCN, three large-scale datasets are used. The results prove that the model can capture spatiotemporal relationships while learning channel-wise topological features effectively, and offer significant improvements and enhancements to the former with comparable computational burdens. The more important is, compared to the previous work,  the new model replies with a multiplicatively smaller scale, and improves the accuracy.

Overall, I suggest accept the manuscript in present form.

Author Response

Thank you for your positive feedback and for taking the time to review our manuscript. We are glad that you found the model and its contributions valuable, and we appreciate your support for its acceptance.

Reviewer 2 Report

Comments and Suggestions for Authors

The research introduces an innovative model, the Attention-Guided and Topology-Enhanced Shift Graph Convolutional Network (AT-Shift-GCN), which significantly improves performance in skeleton-based action recognition by incorporating attention mechanisms and adaptive multi-scale temporal modeling while maintaining computational efficiency. Here are my comments:

While computational efficiency is claimed, the manuscript does not provide concrete evidence such as runtime comparisons, memory usage statistics, or benchmarks on resource-constrained environments.

After introducing the importance of skeleton-based human motion recognition, the introduction directly dives into the vionsion based deep learning, which affects the readability. It is recommended to provide a general background to introduce other deep learning and robotics application first and then narrow down the focus to your topic. For example, the robotics such as Peristaltic transporting device inspired by large intestine structure.

Some figures (Fig 2,3,6,7)are not clearly labeled or referenced in the text, making it difficult to understand the context.

Provide information on where the code and datasets used in this study can be accessed.

Consider expanding the ablation studies to include:

• Different configurations of the attention module (e.g., single-head vs. multi-head attention).

• The impact of removing or altering the adaptive multi-scale temporal convolution.

• Comparison of the performance with and without the ultralight spatiotemporal attention module.

If it is difficult to add the ablation studies, please include a discussion about this.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

Interesting and attractive work. The topic of the work is related to current research trends related to the use of artificial intelligence for calculations and simulations. The authors presented their own computational model, which may have utilitarian applications.

The analysis of the literature review raises concerns about the small geographical diversity of the authors of the works cited in the article. The authors argue basically only with one group of authors of the works. A cursory review of WoS and Scopus shows that this topic is widely discussed by researchers around the world. I am requesting a broader geographical review of the literature.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

This paper proposes and evaluates a new deep learning structure based on shift operators and graph convolutional networks for activity recognition using skeleton information. The main target is to provide good performance with a lightweight model.

The paper is interesting because includes detailed analyses and comparisons to previous works.

Comments to improve the paper:

·       Figures 2 and 3 show the local and non-local shift operations, and they are important to understand the main contribution of the paper. I think these figures require more detail and an expanded explanation.

·       Experimental settings in section 4.1.4. How did you define these settings? Did you consider a validation subset for finetuning?

·       I miss more information about the activities, how many different activities are there in the datasets? How different are they?

·       I’d suggest expanding the description of the performance metrics used in the paper.

·       Can you expand the description of figure 10?

·       Regarding the tables, some descriptions are in capital letters.

·       Tables 5-8. I’d suggest including statistical analyses to see the significance of the results.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 5 Report

Comments and Suggestions for Authors

The manuscript presents an attention-guided adaptive multi-scale temporal convolutional network (AT-Shift-GCN) for skeleton-based action recognition, building upon existing Shift-GCN. The manuscript is technically sound, with claims well-supported by experimental results. The authors provide an analysis of their proposed AT-Shift-GCN model performance on various action recognition tasks from NTU60 RGB+D, NTU120 RGB+D and Northwestern-UCLA datasets, comparing it to the Shift-GCN model. The AT-Shift-GCN proves to be computationally efficient and effective.

Minor:

- Include more results of this study in the abstract section.

- Abstract: “However, most previous works are heavy,….” could be “However, most previous works are resource-heavy,….”

- The writing style and organization are clear but could benefit from more contextualization relative to prior work. 

- Discuss potential developments for future work and limitations of AT-Shift-GCN.

Regards

Comments on the Quality of English Language

There are no apparent English problems.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

The paper has been improved.

Reviewer 4 Report

Comments and Suggestions for Authors

I think the authors have properly addressed my comments.