A DASKL Descriptor via Encoding the Information of Keypoints and a 3D Local Surface for 3D Matching

Round 1

Reviewer 1 Report

 

Paper presents description of a hand-crafted 3D feature description which not only relies on the key point itself but also its statistics deviation from its neighboring adjacent key points. Overall paper is good contribution to the field of 3D point matching and point cloud registration. Paper can benefit from a major revision whereas authors should address the key shortcomings identified below.

 

 

While authors have focused on the manual design of a 3D feature description and tried to justified it by saying that deep learning is data dependent and requires computational power. Such an argument is not generally valid, we have GPU based machines and abundance of labeled and un-labeled data. Authors should provide further justification about not using a deep learning approach e.g., given reference to Edge-devices which may not still be equipped with GPUs or compute resources are very costly and using a manual feature descriptor may be of value if and only if, its inference time is better than the deep learning one — as we are not running any training during deployment.  

 

What does it meant by ‘local frame’, and ‘feature coding’. Without even describing what local frame in fact is, authors have start talking about dividing local frame into local reference frame and local reference axis. 

 

Many sentences in the Introduction section can benefit from English language proof-reading.  

 

The abbreviation of the proposed approach is weird for me i.e., DASKL stands for “deviation angle statistics of keypoints from local points, adjacent keypoints”. Authors are advised to choose a different more descriptive name, possibly without having to write a comma in it. 

 

Other more detailed comments:

There is a typo in abstract, line No. 20. An isolated letter ’t’. 

There should be a space after the period of each sentence. 

There should be a space after the closing brackets of each reference. In some cases, this space is skipped in the manuscript. 

While referring to earlier work, various time the period in “et al.” is missing e.g., “Shah et al”, “Tombari et al”. Please revise and be careful. 

Some sentences are extremely length, please divide them into smaller simpler sentences. For example sentence starting at line 112 and ending at 117. It is not a sentence any more, rather a paragraph in a sentence. 

Sentence at lien 128 is very confusing. 

What is v(p) at line 153. 

Is that a Chinese letter/symbol in equation 2?

What does “Qaccouns” mean?

What are “pairi”, “nimi”, and “pometi” in lines 191 and 192?

Heading of subsection 2.3, should be renamed as “Descriptor Generation”.

Greek letter rho is not shown in figure 1. 

Figure 1 & 3 captions can be enhanced.

Figure 3 and its caption should be on the same page. 

While describing the experimental details; references for the comparative approaches i..e, FPFH, SHOT, and RoS should be provided. Similarly references for the Bologna, and UWAOR and real scene datasets should be provided. The reference to B3R is provided on line 278 but should be provided when it is first listed e.g., on line 270. 

Figure 4 should be consistently on a single page, along with its caption.  

Authors should clarify if the real dataset is the one they gathered? And if so, do they plan to make it publicly available for future research?

Authors should have provided a comparison against a deep learning based method to justify their claim of not using deep learning and focusing on the manual design. 

The compared method are quite old e.g., from 2009 and 2014 — how about comparison against most recent approaches?

 

 

 

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

This manuscript presents an interesting research topic.

Some suggestions:

1) Authors need to thoroughly proofread the manuscript to address typographical and grammatical errors and reword colloquial sentences. E.g., lines 20 to 23 in the abstract , Line 303, Line 307, Equation 11,Lines 426 to 429 and others.

2) The authors can summarily include some metrics from their experiments in both the Abstract and Conclusion to quickly inform potential readers about the findings of the work.

3) A separate related work section can be included in the manuscript, rather than squeezing such details into the introduction.

4) Other than plots, tables can also be used to present the comparisons numerically (e.g., using descriptive statistics) in Section 3.

A very good effort.

Author Response

Dear Reviewer:

Thank you for your decision and constructive comments on my manuscript. We have carefully considered the suggestion of Reviewer and make some changes. We have tried our best to improve and made some changes in the manuscript.

The red part that has been revised according to your comments. Revision notes, point-to-point, are given as follows:

Point 1: Authors need to thoroughly proofread the manuscript to address typographical and grammatical errors and reword colloquial sentences. E.g., lines 20 to 23 in the abstract , Line 303, Line 307, Equation 11,Lines 426 to 429 and others.

 

Response 1: I'm sorry I made a lot of mistakes in the language problem. I have edited the full text in English in MDPI.

Point 2: The authors can summarily include some metrics from their experiments in both the Abstract and Conclusion to quickly inform potential readers about the findings of the work.

Other than plots, tables can also be used to present the comparisons numerically (e.g., using descriptive statistics) in Section 3.

 

Response 2: I am sorry that this part was not clear in the original manuscript, we summarize the indicators in the experiment in the conclusion. The index of the experimental part of this paper is the RPC curve composed of precision and recall. Looking at the accuracy and recall rate alone, it is difficult to explain the performance of the descriptor. For example, when the 1-precision is 0 and the recall rate is low, that is, the descriptor has no error matching, but can not get all the correct matching logarithms of the corresponding features. Accordingly, when the recall rate is very high and the precision is low, it is considered that although the descriptor can fully encode the features of the target to be identified and the scene, and get most of the correct matching, it also introduces a large number of mismatches. Therefore, for a good feature descriptor, its RPC curve should be concentrated in the upper left corner of the graph, that is, the feature descriptor has high recall and accuracy at the same time.

Therefore, the evaluation of the descriptor usually uses the RPC curve directly to show the performance, and the experimental results are given in the summary or conclusion.

 

Point 3: A separate related work section can be included in the manuscript, rather than squeezing such details into the introduction.

 

Response 3: In fact, this is also my idea to separate the introduction from the related work. The structure of the first draft of this article includes a separate part of the relevant work, but the introduction and related work of the MDPI template are in one part, and I modified it according to its template.

 

Thanks again for your decision and constructive comments on my manuscript. We look forward to hearing from you regarding our submission. We would be glad to respond to any further questions and comments that you may have.

 

Author Response File: Author Response.pdf