Scale-Adaptive KCF Mixed with Deep Feature for Pedestrian Tracking

Round 1

Reviewer 1 Report

This paper proposed Scale-adaptive KCF Mixed with Deep Feature for Pedestrian 
Tracking.

 

The paper is well organized and has technical correctness.

It is easier to understand by analyzing and showing the experimental result in detail 

   1) to show example  results (images) for table 2.

   2) to explain advantage/disadvantage of algorithms in table 3

 

 

 

  

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

1. The abbreviations must be defined the first time they appear.
2. The results shown in Table 3 must be discussed in detail to show the significance and effectiveness of the proposed method. For example, the discussions on the following points should be added:
- the reasons for that some conventional methods (KCF, ASLA, and TLD) totally failed to the tracking and those for the dramatic improvement by the proposed method.
- the reasons for the significant differences between SiamRPN and the proposed methods for human6 video and no difference between them for woman and girl2 videos.
- no large differences between the DASiamRPN and proposed methods for all videos.
3. Comparison with respect to computational complexity should be added.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

In this paper, a pedestrian tracking via a novel computer vision technique (a Scale-adaptive Kernel Correlation Filter Mixed with Deep Feature, SKCFMDF) is presented and tested. The proposed method shows a better accuracy than previous methods. Even if the paper is well organized and the topic is of interest for the audience of Electronics, it needs many improvements before the publication. In my opinion, this paper is not ready for the publication on Electronics and the Authors must address the the following general and specific comments.

GENERAL COMMENTS

• First of all, a general improvement of the English is needed, as the paper is sometimes difficult to read.
• L 116: please explain why the image is divided into a 52x52 grid for the smallest detection scale and not into a different number; is this independent form the image size or not? Also for the other scales.
• Table 1 and related comments in the text: in this case, also, the Authors should explain whether or not the sizes are independent from the starting picture size.
• Section 2.4.: the explanation on how the proposed method is able to track pedestrians after an occlusion is too poor, please add a more detailed explanation and an actual example on some images. Moreover, please quantify “long-term occlusion problem”: for how long (number of images) the target can be hidden and still be recognized when it reappears?
• Section 4.2: the Authors state that the best value for the threshold is 0.4, but there is not any general justification of this. Is this value universal of on which parameters it depends? Please explain in a deeper way.
• Section 4.3: the Authors state that “The judgment criterion is that the correct tracking is when the intersection ratio predicted by the model is 0.5” but I haven’t found any justification or explanation of this choice. Please clarify.
• The Authors also highlighted the weaknesses of the method: have they tested it on images of crowdy streets with many pedestrians overlapping?
• Conclusions should highlight the main results, not only the future potential development.

SPECIFIC COMMENTS

• Please do not use acronym in the abstract.
• The references should be numbered in the order they are mentioned in the text.
• Line 30: to help non specialist audience, please explain the acronyms (RPN) the first time they are mentioned in the main text.
• L 34: “However, FOR the SiamRPN model is still difficult to…”
• L 37: please explain the acronyms (SIFT) the first time they are mentioned in the main text.
• L 40: please explain the acronyms (KCF) the first time they are mentioned in the main text.
• L 46: please explain the acronyms (HOG) the first time they are mentioned in the main text.
• L 68: please add a reference for the PASCAL VOC data set.
• L 72-79 are a repetition of lines 43-51: please remove or rephrase.
• L 131: please replace the comma after “covering the target” with a full stop.
• L 132-134: please rephrase the sentence “At this time,…” as the main verb is missing.
• L 152: L 46: please explain the acronyms (ELU) the first time they are mentioned in the main text.
• L 156-157: please rephrase the sentence “Using neural network …” as the main verb is missing.
• L 177: “It” should be lowercase.
• L 233: please explain the acronyms (FHOG) the first time they are mentioned in the main text.
• L 270-271: please rephrase the sentence (there is not any subject).
• L 273: please add a space after the full stop.
• Figure 6: please provide a higher resolution image.
• L 295: please rephrase and correct the sentence “The σ value Set to 0.4.”.
• L 307: please correct “human6” (in table 3 as well).
• L 308: please correct “girl2” (in table 3 as well).
• L 310: please explain the acronyms the first time they are mentioned in the main text.
• L 324: please add a full stop after “[18]”.

In conclusion, I propose the paper for major revisions.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

None

Reviewer 3 Report

In my opinion, this revised version of the paper is much better than the previous one. All of my points have been fixed, so I think the paper can be published in this version.