EnsembleVehicleDet: Detection of Faraway Vehicles with Real-Time Consideration

Round 1

Reviewer 1 Report

Brief summary

The manuscript deals with the problem of the detecting of surrounding vehicles in autonomous driving using deep learning method. The main task of the work is to increase the efficiency of detecting the far vehicles. To carry out this task authors propose the ensemble algorithm which key idea is to divide the picture into two sections. Both sections are then processed separately and the results are finally merged. The experiments conducted on Agroverse data set confirmed high efficiency of the developed algorithm.

Broad comments

The term “small vehicles” used in the manuscript is confusing. It suggests the small size cars while in the manuscript authors mean the cars located in the larger distance. I suggest changing the title replacing ‘small’ with ‘far’ or ‘distant’. ‘Small’ can be used in the manuscript body but authors must give a proper explanation in the Introduction.

There is only a general information about the performance of the EnsembleVehicleDet method in the Conclusions section. It should be extended by information about the performance comparison of various variants of Model B.

Specific comments and minor mistakes

Authors are overusing referencing. It is not necessary to give the reference in every place the term is used (Agroverse,YOLOv7).

Figure 3: The original image without RoI marking would be helpful

Figure 4: Authors are asked for some comments explaining what the visible shapes are connected with.

Figures 8-10: The pictures have insufficient resolution. The details showing processing results are hardly visible.

Algorithms pseudocode

Presenting algorithms by pseudocode is a good idea but the authors should take some more effort in increasing their legibility. It would be helpful to introduce comment lines describing the functions of the particular pseudocode sections.

As the code blocks are not distinguished by the brackets ‘{ }’ or ‘begin – end’ it is essential to keep the proper indents.

 

Algorithm 1:

 

The fragment

AccumulateImage_S.rows = rows from AccumulateImage_S

AccumulateImage_S.cols = columns from AccumulateImage_S

Frame.rows = rows from Frame

Frame.cols = columns from Frame

is not clear. Does it mean that AccumulateImage_S and Frame retrieve their dimensions from themselves?

 

The lines:

X_Range_1 = AccumulateImage_S.rows / 2 - AccumulateImage_S.rows / 4

X_Range_2 = AccumulateImage_S.rows / 2 + AccumulateImage_S.rows / 4

can be written as:

X_Range_1 = AccumulateImage_S.rows / 4

X_Range_2 = AccumulateImage_S.rows * 3 / 4

 

The next fragment indents are wrong. I am guessing they shold look like this.

for i = 0 to (AccumulateImage_S.cols - 160) do

sum = 0

for x = X_Range_1 to X_Range_2 do

for y = i to (i + 160) do

sum += AccumulateImage_S[Frame.cols * y * 3 + x * 3]

sum += AccumulateImage_S[Frame.cols * y * 3 + x * 3 + 1]

sum += AccumulateImage_S[Frame.cols * y * 3 + x * 3 + 2]

if Max_sum < sum then

Max_sum = sum

RoI_cols = i

Please check them.

 

Algorithm 3:

Some code fragments have wrong indents. Please correct them.

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

The authors propose an ensemble method that can accurately detect small vehicles while meeting real-time requirements.To improve the article even more, my suggestions are as follows:

1.The novelty of the paper needs to be further strengthened.

2.The authors have omitted many important studies related to the current state of research on deep learning models. In the introduction, it is clear that the authors do not have enough information on the existing research and suggest that the authors must examine more research done by others. See for example “PM2.5 volatility prediction by XGBoost-MLP based on GARCH models”and"Haze Risk Assessment Based on Improved PCA-MEE and ISPO-LightGBM Model".

3.The author lacks a description of the dataset section and suggests adding this section.

4.Lines 191-192 ,253-254, 269-270 are suggested for the appendix.

5.As shown in Table 3, the higher the accuracy rate, the lower the recall rate instead, can you explain the reason for this situation?

6.300-308 should be placed before Table 3, and Table 4 should also be placed after rows 312-331.

7.The figure is the same, explained first, then shown.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

1. Authors should add the more caption about the accuracy (AP0.5, AP 0.5-S, AP 0.5-M and AP 0.5-I). What the meaning of these accuracy ?

2. What the definition of Model A and Model B in Figure 2? Authord should descript it at the first paragraph of  Section 3 and Figure 2. 

3. What meaning of different color marsk in Figure 3?

4. Authors should add the defintion of performance index (TP, FP, Recall etc.) in Table 3 in the study. 

5. Is it possible to show the ROC and AUC to verify the performance of proposed method ? 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

The authors meticulously considered all my comments. In my opinion, the manuscript can be accepted for publication.

Reviewer 2 Report

The author has made changes in line with my comments. Good luck!

Reviewer 3 Report

Authors revised the draft according the reviewer's comments.