Deep Hybrid Compression Network for Lidar Point Cloud Classification and Segmentation

Round 1

Reviewer 1 Report

In this study, authors introduced a novel hybrid compression method  improves model accuracy and reduces model memory consumption of LiDAR point clouds processing  based on relaxed mixed-precision quantization, relaxed weights pruning, and knowledge distillation. This research has a solid foundation of mathematics and computer, but the manuscript was not suitable for publishing at current version.

1. Introduction section was not force on the detail problem this study want to solve. It should be more attention on this issue.

2. The information of data used in this study was not described in the manuscript.

3.  Too much detail about the algorithm described in the method section. I think most of them can be moved to appendix.

4. The logical  expression of the method part is poor. And the core of the algorithm was discrete.

5. Line 313 to Line327 was the parameters setting and it should be moved to method sections.

6. The results section, there were only some simulated results but there are few comparison with other research. It's a results about talking to yourself.

7. Discussion was thin and more like a repeat of the results. And this section should be rewritten carefully and deeply. And should focus on the comparison with other research and disadvantage, advantage and uncertainty analysis should be added.

8. There also have some error in formula serial numbe and section serial numbe.

9. The detail of accuracy assessment in method section was missed.

 

special comments

 

1. “[16] build a framework to solve the constrained optimization problem efficiently by using ADMM. [17] introduces a Hardware-Aware Automated Quantization (HAQ) framework which leverages reinforcement learning to automatically determine the quantization policy. [18] propose a hierarchical DRL based kernel-wise network quantization technique, AutoQ, to...” 

This is not a format of citation for a scientific paper. It should be described as “Author do what [reference]”, please revise all incorrect citation format.

2. Wrong formula number,  there are two equation (3) between line 197-198 and line 203-204; two equation (6) between line 214-215 to line225-226.

3. Wrong section number. There are two section 3.1.2 at line352 and line 365.

Author Response

Thanks for your valuable comments. I appreciate and benefit a lot.  Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

This paper proposes a deep hybrid compression method based on relaxed mixed-precision quantization, relaxed weights pruning, and knowledge distillation for LiDAR point cloud classification and Segmentation.

Please add a general workflow and describe all phases. Explain the used data and input to the model.

Please clearly indicate the benefits of your model compared to the backbone algorithms.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

This manuscript presents a deep hybrid compression network based on relaxed mixed-precision quantization and mixed-pruning to enhance model accuracy and reduce model memory consumption compared with a uniform quantization method. The proposed method is tested and evaluated with other backbone networks which utilize full precision and uniform quantization, and the results show the proposed network is competitive in terms of accuracy and model size.

The proposed approach has merit and the standard is high, therefore, I recommend its publication after minor revision.

I suggest the final check of the manuscript format. Table 5 is a little truncated.

This manuscript is well-written, however, I suggest a final check to correct minor typos.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 4 Report

Dear authors,

your study is well-organised and is a hot topic for LiDAR data science.

Compression accuracy have been able to increased. Good work.

Congratulations

Quality of english is good.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 5 Report

The paper presents a novel deep hybrid compression network for Lidar point cloud classification and segmenation. The paper is weill organized and lots of experiments are conducted to demonstrate the effectiveness of the proposed method. However, the following issues need to be considered:

1.        The novelties of the proposed method are not clear. What are the limitations of current methods? How do the proposed modules can solve these limitations?

2.        The main contributions should be rewritten. Usually the expression “a different xxx” cannot demonstrate the novelty of the proposed method.

3.        Comparative experiments of algorithms need to be strengthened. For example, the comparisons of different methods in Figure7 and Figure 8 could be illustrated or explained further.

 

 

1.        The overall writing needs further polished.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 5 Report

No more comments.