Sparse 3D Point Cloud Parallel Multi-Scale Feature Extraction and Dense Reconstruction with Multi-Headed Attentional Upsampling

Round 1

Reviewer 1 Report

The problem statement is not clear. Author should brielfy mention the problem statement in the Abstract section and in detail in the Introduction section. Also mention that the proposed solution will solve what kind of issues and where the proposed solutions can be used i.e. applicability of the work in real world.

In related work section, please specify the research gaps and address the loop holes in previous studies and then suggest your solution as a novel proposal to solve such problems.

You may use graphs to illustrate your results. Did you compare other exisiting approches with the proposed approach on diiferent datasets? If not then I recommmend you to do that.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

The author introduced a point cloud upsampling method by combining a parallel multi-scale feature extraction module and a multi-head self-attention module. The performance of the paper is impressive. 

Here are some questions. and suggestions to improve the paper.

1. Please phase the expression of the introduction session. "The ultimate embodiment of computer vision is 3D vision, and the expression of 3D vision is point clouds." The statement needs evidence. 

2. Multi-head self-attention module is an important technique in this paper, it will be great if the author can add more experiments on the influence of head number. Will the model be benefited from adding more heads? 

3. Will adding more Multi-head self-attention blocks be beneficial like other transformer-based models?

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

This paper proposes a GCN-based point cloud upsampling method with parallel multiscale feature extraction and multihead self-attention. Experimental results demonstrate the effectiveness of the proposed method, compared to several state-of-the-arts. Some detailed comments can be found as follows:

1. The differences between the proposed method and PU-GCN should be pointed out in the introduction. Even the PU-GCN is not described in detail, but it is also a GCN-based point cloud upsampling model which is rather relevant to the proposed method.

2. Removing PMS or MHA and using single attention/feature are suggested to be verified in the experiments separately.

3. To provide better understanding for readers, some GCN-based image processing works are suggested to be reviewed, including Rain streak removal via dual graph convolutional network, Blind omnidirectional image quality assessment with viewport oriented graph convolutional networks, Adaptive hypergraph convolutional network for no-reference 360-degree image quality assessment, etc.

4. In the ablation study, using PU-GCN for both PMS and MHA is confused. Does the proposed framework removing PMS or MHA be equal to PU-GCN? Please clarify this point.

5. It is suggested to further proofread the paper. The descriptions should be unified, such as PointNet and pointnet.

 

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The revised version looks well. However, the quality of English is not good and paper is difficult to understand. I suggest authors to improve the quality of English. The authors may use MDPI English editing services or some other services for English and grammar editing. The article can be accepted after revising the English and Grammar.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

I have no further comments.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 3

Reviewer 1 Report

Accept

Author Response

Thank you for your approval.