AD-SiamRPN: Anti-Deformation Object Tracking via an Improved Siamese Region Proposal Network on Hyperspectral Videos

Round 1

Reviewer 1 Report

The paper introduces a hyperspectral video tracking algorithm based on an improved Siamese region proposal network. The proposed model focuses on the deformation problem in the field of tracking, which is proved by adequate experimental results. Overall, the manuscript includes a very interesting topic and it is well structured and organized. I believe that this manuscript could be accepted after the following small suggestions.

Q1. Please check the grammar and spelling.

For example:

On the first page, ‘pf’ should be changed to ‘of’.

Below Eq.8, ‘image’ should be changed to ‘images’.

On page 15, ‘our follow work’ can be revised to ‘our following work’.

Q2. Check all figures to make the font size appropriate.

Q3. It is essential to include directions for future work. For example, this paper uses the band selection method to obtain the network input. Yet, maybe you can design a novel network with inputting the raw hyperspectral image in the future.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

This paper proposes an anti-deformation hyperspectral video tracking method via the genetic optimization and template accumulation based SiamRPN model. There are mainly three modules, including the band selection, transfer learning and template updating. Experiments, including the ablation experiment, have been conducted to evaluate the effectiveness of the proposed method. If following problems are solved, this manuscript can be accepted.

(1) Some spelling mistakes and description should be corrected.

1. In Figure 5, the "Np-Update" needs to be changed to ‘No-update’.

2. In the last paragraph of the second page, the word ‘tha’ in ‘For the reason tha the UpdateNet...’ should be ‘that’, ‘The the effect of the TL strategy on...’ requires the removal of ‘the’.

3. Please change φ to ψ in Figure 4

4. Please add a reference to the HOT2022 data in Section 3.1.

5. Please provide a description of the video type false-color in Table 3.

(2) In general, both hyperspectral data processing and the deep network based method are time-consuming. Thus, how to ensure the efficiency of algorithm, and what is the speed of this tracking algorithm?

(3) It is suggested to add some references about application of hyperspectral images in other fields.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

This work presents a novel method for tracking objects in hyperspectral video. The method is based on the used a pre-trained SiamRPN model combined with transfer learning and spectral template model and band selection. The results obtained are competitive with the best methods available. In general I find this study very interesting and complete, but I have some suggestions for a revised version:

- There are not many details in sections 2.1, 2.2, 2.3 and 2.4. Some papers probably add excessive details when explaining the fundamentals of the method if those are well known, but I find that more details on the method will help the readers in this case. In particular, there are no details on the architecture of the networks used. It is not easy to understand for instance what is the novelty of HSUpdateNet with respect to UpdateNet. Or why there are improvements with respect to UpdateNet as the authors write at the end of section 2.4.

- Related to the previous point, there is no explanation of the parameter sigma in the definition of smooth_L1.

-I do not find an explanation of why 6 videos are selected. I guess this is from the 35 test sets in HOT2022. Are they selected to illustrate the method in the article? Is there a criterion to select them?

- The video descriptions in page 8 for Fig. 6b, 6c, 6d, 6e, 6f do not match the images and do not match the description of the video tracking results later in the text. 

- In Table 7, column DT@20P, it is highlighted in blue color BS-SiamRPN when actually SiamRPN++ has a higher value of 0.971 equal to the paper method

- It is argued that the optimisation algorithm speeds up the tasking task, buty there are no vlaues of time/speed of the method or the different methods. It is reasonable, but this conclussion shall be supported by some evidence. It would also be interesting to have some values of the speed of the method. 

- There are some sentences and expressions in the text taht sound extrange at least to me (e.g. in page 4, "intrinsic information of HSVs should be got through feadure..." or "...correlation coefficients in Fig. 2 insists that..." or "...entropy loss, λ is a hyperparameter the two loss"). There are also some typos. I recommend a thoroughful re-reading of the text to the authors.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

This study proposes an anti-deformation object tracking method via improved siamese region proposal network on hyperspectral videos. Its main contribution consists in using the SiamRPN model and apply it to hyperspectral videos object tracking.

 

The document is sometimes hard to read and follow.

The English needs major spell checking.

The document is well supported with references.

The subject of the paper has great potential of application.

 

 

The proposed work main weakness is the lack of clear comparison of the proposed method with other state of the art methods. Also, the authors missed the opportunity of comparing the results with the other presented feature dimension reduction methods.

 

The Abstract is constituted of disconnected sentences often not clear. Authors should deeply revise the Abstract pointing out clearly the motivations, the main contributions, the general structure of the proposed methodology and obtained results.

 

Authors should support with bibliographic references the following sentence presented in the Introduction section “… it is limited in describing physical characteristics of the image, which leads to tracking failure, as the target is deformed.”

 

Subsection 2.1 should include a diagram representing the SiamRPN model.

 

In subsection 2.2 authors should improve the reasoning related to the selection of feature dimension reduction method.

 

In results section authors should clearly separate and compare the performance of the presented method with the other methods in tests where slow motion vs fast motion occurs in moving objects, large deformation vs small deformation of objects.

 

Authors should also clearly present the behavior of the presented and other tracking methods when in presence of occlusion.

 

 

Finally, authors should address the reasoning behind using single object tracking.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 4 Report

Since the main issues pointed out in the previous review were addressed by the authors i advise the manuscript to be considered for publication.

Author Response

We are sincerely grateful to your insightful comments and thoughtful suggestions again on our manuscript.

We carefully check the content of the whole manuscript again.

Thanks again for your comments.