Adversarial Examples in Visual Object Tracking in Satellite Videos: Cross-Frame Momentum Accumulation for Adversarial Examples Generation

Round 1

Reviewer 1 Report

The paper addresses an important and modern topic of object tracking in videos using trained neural networks. The paper is written well enough. However, it can be improved. In this sense, the following can be done: 

1) Is it possible to illustrate adversarial example?

2) I don’t like the phrases “[23] investigated…” or “[25] attacks..” Besides, write such type of sentences in the same tense.

3) Maybe, the paper https://www.researchgate.net/publication/343414601_Efficient_Adversarial_Attacks_for_Visual_Object_Tracking is worth analyzing and citing?

4) Please give a reference or references where Attack Success Rate and Attack Precision are used for solving similar tasks.

5) Improve the plots in Fig. 5

6) What is the difference between remote sensing images and other types of images for which object tracking in videos has to be carried out?

Author Response

We appreciate for Reviewer’s warm work earnestly, and hope that the correction will meet with approval. We will upload feedback comments.
Once again, thank you very much for your comments and suggestions.

Author Response File: Author Response.pdf

Reviewer 2 Report

No comments. A very nice work well done and documented

Author Response

We appreciate for Reviewer’s warm work earnestly, and hope that the correction will meet with approval. 
Once again, thank you very much for your comments and suggestions.

Reviewer 3 Report

This paper propose an adversarial example attack in the field of RSI visual object tracking. Experiments in SatSOT and VISO verify that the proposed method can effectively attack Siamese trackers. 

The writting and structure of this manuscript is good, However, there are some concerns as following:

1. It was only conducted on SatSOT and VISO datasets, too few datasets. The author had better increase the number of datasets, such as UAV123, DTB. 

2. Only the classical adversarial example attack algorithm is introduced, and the adversarial example [1] and siamese trackers [2] in recent years is lacking.

[1] Few-Shot Backdoor Attacks on Visual Object Tracking, ICLR 2022

[2] Reinforced Similarity Learning: Siamese Relation Networks for Robust Object Tracking, ACM MM 2020

3. The current manuscript needs to be further polished in terms of grammar and formats. There are some errors, such as:

[1] Visual object tracking（VOT） // The success of visual object tracking (VOT) 

[2] The line spacing of pseudo-code algorithm 1 is too large, which affects the appearance.

[3] 4.1.1. APE  //  4.1.2. AOR  //  4.1.3. Attack Success rate and attack Precision

......

4. For the same calculation formula, α is used in formula (9) and δ is used in line 5 of the pseudo-code.

5. There are two formula (7) and two Figure 5, which are unclear.

6. Too much space is devoted to evaluation criteria.

7. The text in Figure 5 on page 14 is too small and distorted when enlarged. And this outline boarder is imcomplete.

8. The attack method just focus on SiamRPN series, so what about anchor-free siamese trackers (such as SiamCAR, Ocean ...)

Moderate editing of English language required

Author Response

We appreciate for Reviewer’s warm work earnestly, and hope that the correction will meet with approval. We will upload feedback comments
Once again, thank you very much for your comments and suggestions.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

I am satisfgied by the answers and corrections done. 

Reviewer 3 Report

The authors have addressed most of my concerns, so this version can be accepted after checking and revising the text writting. such as 

[1] where M is the proposLal number

[2] formula (5) can be improved 

please check the text writting.