WeaveNet: Solution for Variable Input Sparsity Depth Completion

Round 1

Reviewer 1 Report

Dear authors;

Your paper entitled “WeaveNet: Solution for Variable Input Sparsity Depth

Completion” I have several comments that I wish to be useful for you:

1- The abstract needs more interest and rewriting some paragraphs.

2- There are still some aspects that can be improved (for grammar and punctuations). Improve the technical writing of your paper, where there are several grammatical errors and spelling I think they need to be checked out.

3- The conclusion needs more efforts to elaborate the achieved results with respect to the future work,

4- The practical part is very important,

5- Future work is an important part of the conclusion.

I loved this work and I feel it is very good. I hope these comments would help you improve this work after a major revision.

Regards

Author Response

Dear Reviewer,

Thank you for your insightful remarks.

1. I have rewritten the abstract to put more emphasis on the robustness against input density as a feature of WeaveNet.

2. Thank you for pointing that out. I found some mistakes and corrected them.

3. I have rewritten the conclusions chapter. I explained why imaging radars are likely to benefit from a solution similar to WeaveNet and noted the ease of transfer learning on a different sensor with WeaveNet.

4,5. Thank you for noting that, I put additional emphasis in the introduction on the need for robustness of an automotive-grade depth completion solution to weather conditions that might affect the input density and elaborated more on the future work in the conclusion part.

Best regards,

Mariusz K. Nowak

Reviewer 2 Report

In this paper, a Lidar analysis method is proposed for depth completion which could be essential for auto-driving or surveying. The overall framework is described in detail but no mathematic formulas or processing algorithms are supplied, which makes the paper weak in the theoretical aspect. Also in the experimental part, the results indicate that the proposed method is not out-perform the state of art work. Then what the the value of the proposed method should be further discussed. For example the authors should give an analysis on computation complexity or some other benefit that can be achieved by the method. 

Some other specific suggestions:

In section 4, Only one scene is given. More results in different scenarios should be supplied.

Author Response

Dear Reviewer,

Thank you for your insightful remarks.

The main goal of the paper was to present a solution that is robust to different levels of input sparsity (in real life scenario such differences may arise for example due to changing weather conditions – there are significantly fewer LIDAR measurements in heavy fog). Unlike other solutions present in literature, WeaveNet achieves it by succesfully utilizing one set of weights over 2 orders of magnitude of input density (from 200 to 20,000 measurements in the FOV). After your remarks I put more emphasis on this robustness feature in the abstract, introduction and conclusions. Additionally, I added information on the ease of transfer learning between different sensors using WeaveNet architecture and presented the results obtained by using network trained on KITTI (Velodyne LIDAR, 64 channels) dataset and applying it to data from Pandora LIDAR (40 channels). It is shown qualitatively in Figure 9. I think adding the Figure 9. resolves your specific remark about only one scene in section 4.

Best Regards,

Mariusz K. Nowak

Reviewer 3 Report

• Paper name: WeaveNet: Solution for Variable Input Sparsity Depth Completion, by author Mariusz Karol Nowak

  The paper proposes method to enhance the data provided by 360˚ LiDAR sensors used in automotive industry for automatic detection of the object placed near the vehicle. The method operates with both LiDAR and cameras information.

  • Broad comments highlighting areas of strength and weakness. These comments should be specific enough for authors to be able to respond.

  The paper is well structured, containing four parts:

  Chapter 1. Introduction with 7 references compressed in one paragraph.;

  Chapter 2. Related work – this is a good review of the main papers in the field of image processing;

  Chapter 3. Our method! Covering the Methods and materials…;

  Chapter 4. Experiments;

  Chapter 4! Conclusions, are well developed and are using the findings of the paper.

  For the reviewer point of view, the paper has plenty of knowledge and hard work. Even so, some minor improvements can be made:

  1.     The title and the abstract must avoid dedicated abbreviations or commercial names and they must challenge the reader enough in order to read the paper. It is not the case here. Please, add in the introduction chapter all the knowledge you think it can be inserted, all the particular terms you will use and the references.

  2.     The paper is too dense, being difficult to follow the demonstration, especially when you must jump all over though the references. References are not clear and, even if you are reading the whole paper, is not possible to fully understand what information was used from the referenced paper.

  3.     Some abbreviations are missing (ReLU).

  4.     Who can use the findings of your paper?

  5.     Can you apply this method for data provided with another device? As far as I know, a year ago, Velodyne HDL-64E did steps into the history.

  6.     What is the relation of this paper with Electronics? It’s mostly Video and Image Processing.

  7. Also, consider changing the name WeaveNet – was used, on a similar paper, in 2017: https://arxiv.org/abs/1712.03149

  • Specific comments referring to line numbers, tables or figures. Reviewers need not comment on formatting issues that do not obscure the meaning of the paper, as these will be addressed by editors.

   The papaer looks fine, maybe the editors will ask for clearer images.

Comments for author File: Comments.pdf

Author Response

Dear Reviewer,

Thank you for your insightful remarks.

1. Thank you for reminding me about that. I removed the Velodyne name from the abstract. Additionally I have rewritten the abstract and introduction to put emphasis on the main selling point of WeaveNet – ability to perform well under very different input densities using one set of weights. I also pointed out when such conditions may arise in real life (e.g. in heavy fog).

2. I went through the paper once again and in fact I found some references that were hard to disentangle. I made some changes, and hopefully it is easier to read now.

3. Thank you, I fixed that.

4. I added more specific information about usefulness of WeaveNet in 3 cases:

- Robustness to changing weather conditions

- More explanation on why imaging radars might benefit from WeaveNet-like approach

- I have shown ease of transfer learning with WeaveNet-like solution by showing that network trained on Velodyne 64 performs qualitatively well on Pandora (40 channels) data (Figure 9.)

5. I have shown performance on Pandora (Figure 9.)

6.While the paper itself deals more with software, rather than hardware layer, I think it is relevant to the ‘sensing’ and ‘data fusion’ keywords of the Special Issue "IoT for Intelligent Transportation Systems"

Best Regards,

Mariusz K. Nowak

Round 2

Reviewer 1 Report

I have no further comments

Author Response

Thank you for your review.

Best regards,

Mariusz K. Nowak

Reviewer 2 Report

The authors have revised the paper according to the comments. Also some new results are included to explain the feature of the the proposed method. I think this paper could be considered for publication after some minor revisions on language checking and background optimization.

Author Response

Dear Reviewer, 

I found and corrected some small language errors. Additionally I added an image containing a schematic representation of a residual block used in my network to make the research more reproducible.

Thank you for your review.

Best regards,

Mariusz K. Nowak

Reviewer 3 Report

I would like to thank the author for the revised version of the paper.

Although there are some improvements, the overall scientific soundness is still low for an article. The presentation is not enough to replicate the  experiment presented in the article. 

Some issues about the commercial name - as been used before by WAEVEWORKS company.

Author Response

Dear Reviewer, 

I found and corrected some small language errors. Additionally I added an image containing a schematic representation of a residual block used in my network to make the research more reproducible.

When it comes to the Weaveworks company, they do have a cloud networking toolkit called Weave Net, however it is spelled differently than my solution (WeaveNet). Given that 'weave' is a common word and my solution deals with problems completely unrelated to cloud computing, I don't think it is a copyright infringement.

Thank you for your review.

Best regards,

Mariusz K. Nowak