Temporal Context Modeling Network with Local-Global Complementary Architecture for Temporal Proposal Generation

Round 1

Reviewer 1 Report

--A very solid paper with convincing results.

--There is no discussion on computational complexity of your method.  Is computational efficiency of no concern in this application?  A brief discussion should improve the paper.

-- Minor grammar/sentence errors are found throughout and must be corrected.  As examples, corrections and suggestions for the first few pages are attached.  

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

This paper proposed a temporal action proposal generation method utilizing a temporal context modeling network. The method contains a base network with dilated convolutions for addressing the varied temporal length of actions, followed by an action completeness module and temporal boundary generator. The model was evaluated on two popular benchmarks, and the proposed network yields better performance compared to the state-of-the-art methods. 

The proposed method is interesting and effective in addressing temporal action proposal generation. I have the following comments regarding the paper.

1. The authors should show how the “delta” parameter is chosen for feature encoding introduced in line 165. How does the performance behave with various “delta” values?
2. The authors are encouraged to include some failure cases and the probable reasons for the failures.
3. In line 331, remove the sentence “we are sorry about this.” How did the authors re-implement the methods? Are the official codebases already available, and did they not reproduce the results provided in the papers? The authors should clarify the details. If official codebases are not provided, why the authors did not report the results from the paper?
4. In line 323, the authors are encouraged to cite relevant papers that introduce the evaluation metrics. It is unclear from the paper how the evaluation metrics are computed.
5. The proposed method lacks details in evaluation. The authors are encouraged to show a detailed analysis of what type of actions the method detects well and what they perform less accurately. How does the method perform on longer temporal actions and how does it on shorter ones? The authors are encouraged to show more detailed evaluation results based on duration to show that it meets their claim in line 66.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

This is a very  interesting and complete work presenting a novel approach for tempral proposal generation. Some comments:

- Highlight all assumptions and limitations of your work.

- Conclusions should provide some lessons learnt.

- Related works section does not mention recent research efforts in close related fields where a type of video sequence has to be treated in non-standard ways including different types of implicit dimensionality reduction techniques. Authors are advised to refer to the following related articles to add some discussions: [1] A predictive hybrid reduced order model based on proper orthogonal decomposition combined with deep learning architectures, Expert Systems with Applications, 2022 [2] Deep Learning-based Action Detection in Untrimmed Videos: A Survey, IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022 [3] Model-free short-term fluid dynamics estimator with a deep 3D-convolutional neural network, Expert Systems with Applications, 2021

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx