A Deep Learning Based Method for Railway Overhead Wire Reconstruction from Airborne LiDAR Data

Round 1

Reviewer 1 Report

This manuscript proposes a workflow for railway overhead wire reconstruction using deep learning for wire identification collaborating with the RANdom SAmple Consensus algorithm for wire reconstruction. In general, the manuscript gives a good description of the proposed methodology, and the data analysis and results are fully discussed. However, the manuscript needs to be improved, and the analysis presented in the manuscript is somehow limited and required to be extended. More comments:

1.     More details are required, such as the TP, FP, TN, and FN in Section 4.3. Please explain specifically what the correct category is and what the wrong category is in the context of the corresponding experiment.

2.     PointNet was proposed in 2017, there are lots methods for point cloud processing techniques in 2022, such as “R-CNN-Based Crack Identification”. Please refer to some new literature to expand the applicability of the method proposed in this manuscript.

3.     Some Figures need to be improved, such as the Figure 1 which was included in the Figure 2. It will be better to show the contribution or motivation of this manuscript in the Figure 1.

4.     Please explain why the four indicators of MIoU, ACC, Fitting rate, Fitting error are selected for evaluation.

5.     There are some weird expressions, such as the title of section 2.3-Learning based Methods. It will be better to use machine learning to describe this section.

Author Response

Dear Reviewer:

The authors would like to thank you very much for giving such valuable comments and suggestions for improving the manuscript. The responses to each specific comment are given in the attached PDF. Meanwhile, the modifications with regard to the comments are also highlighted in blue color in the re-submitted PDF file of the revised version of the manuscript.

Thank you very much.

Best regards,

Jinhu Wang (on behalf of the authors)

 

Author Response File: Author Response.pdf

Reviewer 2 Report

A review of ‘A Deep Learning Based Method for Railway Overhead Wire Reconstruction from Airborne LiDAR Data’ 

The paper proposes an approach for segmenting and fitting railway overhead wires. The approach is based on two steps: In the first step, a modification of PointNet is used for segmenting wires in a 3D point-cloud. In the second step, RANSAC is used for fitting a second order polynomial to the points in each segment. 

In the caption of Fig. 1 ‘colorizes’ -> ‘colorized’ 

P. 3: ‘An algorithm that filtering out’ -> ‘[...] that filters out [...]’

Maybe it could be better to provide references for K-Means, DBSCAN and AGENS on p. 3? 

P. 4 ‘the average completeness [...] are 0.92, 0.99, and 0.91’. Since percentages were used earlier (p. 3), maybe it is better to use percentages here as well for consistency (i.e. 92%, 99%, and 91%) 

Table 1: I don’t see the point of listing the proposed method and giving as characteristics: ‘Extracting the wires of railway using proposed method’. 

P. 6 ‘ground points are downsampling’ -> ‘[...] are downsampled’ 

I am not sure if the image of an octree (Fig. 4) is necessary. This is a well known data-structure (in geometry processing). 

The description in lines 230 - 231 is not very clear. The neighborhood to a point is computed, and PCA is applied to that neighborhood. PCA is not an algorithm for computing eigenvalues (as the text seems to imply). Please try to reformulate these two lines. 

P. 8: It is not clear why A+L+P = 1. (It doesn’t seem correct) 

P. 9, lines 251 - 257 and Fig. 8 are not very clear. The k-nearest neighbor search is done on points in dimension 16? How do you arrive at the tensor of dimension Nx64 from a Nx3 tensor and Nx16 tensors (how many of them are there?) 

In Fig. 9: It is not clear how the global features are used in the PointNet module (it is a 1x1024 vector, how is it combined/where is it used in the Nx1984 final tensor?)

The section on the computation of the original wire span is brief. Could you please provide more details. 

It is not clear to me why the curves should be in the XOZ plane. These should be 3D curves. I assume that this is a simplifying hypothesis. 

In Section 3.3 RANSAC is used for fitting a curve, was least-square fitting insufficient? 

Section 4.1 ‘Data description’ (typo) 

The index ‘k’ does not appear anywhere in the summand in (9) 

In (12) how do you define the distance? Do you use the ‘orthogonal’ distance to the curve? Or simply the ‘vertical’ distance? 

The section 4.6.1 is not well written. Please try to reformulate. What is a ‘gradient explosion’? 

P. 17: ‘the effectiveness of the module proposed’ -> ‘proposed module’ 

P. 20: ‘98.52% and 0.8994’ -> Use percentages everywhere 

Line 485: ‘the MLP are used in concatenate’. Please reformulate. 

Ref. 3: Please check the names (The first names are shown instead of the family names) 

Author Response

Dear Reviewer:

The authors would like to thank you so much for giving such important and valuable comments for us to improve the manuscript.

The responses to each specific comment are given in the attached PDF file. Meanwhile, the modifications with respect to the comments are also highlighted in orange color in the re-submitted PDF file of the revised version of the manuscript.

Thank you so much again.

Best regards,

Jinhu Wang (on behalf of the authors)

 

Author Response File: Author Response.pdf