Low-Rank Multi-Channel Features for Robust Visual Object Tracking
Round 1
Reviewer 1 Report
In this manuscript, a novel visual object tracking method based on kernel correlation filters (KCF) has been proposed. The proposed method has incorporated rich features and reduced the rank of feature dimension to boost the tracking performance. The proposed method has been evaluated on a publicly available benchmark dataset and has been compared against several state-of-the-art tracking methods. The experimental results have shown the proposed method can achieve promising results with fast tracking speed.
In general, the manuscript is well-organized and the techniques used in the current manuscript are reasonable. However, the manuscript still needs improvements to be a publication.
The comments of the reviewer are as follows.
The author should proofread the manuscript carefully. There are some minor grammar or typos, e.g., On page 1, “The Generative models” should be “The generative models”, On page 2, “Discriminative tracker largely depends” should be “A discriminative tracker largely depends”, On page 2, “The benchmark datasets [9] is” should be “The benchmark dataset [9] is”, On page 2, “Section III contain” should be “Section III contains”, On page 2, “The discriminative model in [19] and [20], employ” should be “The discriminative models in [19] and [20], employ”, On page 2, “Sparsity-based Collaborative Model (SCM) [54] and Adaptive Structural Local-sparse Appearance model (ASLA) [57] is proposed” should be “Sparsity-based Collaborative Model (SCM) [54] and Adaptive Structural Local-sparse Appearance model (ASLA) [57] are proposed” On page 3, “The proposed tracker incorporate the” should be “The proposed tracker incorporates the”. Some references are duplicated. For instance, Reference [30] and Reference [53] are the same, and Reference [8] and Reference [49] are the same. The author should use a different font to highlight the best performance in each table. Most tracking methods used for comparison are out of date. If the author can compare the proposed method against most recent methods it would be more convincing.
Author Response
Dear Reviewer,
First of all, we would like to sincerely thank you for their thoughtful comments, questions, and suggestions. We hope that we satisfactorily respond to each one of them.
Attached files are as below
Revised manuscript.
Revised manuscript with marks for the updated sentences.
Reply to the Comment of the reviewer.
Thank you for your contributions to this manuscript,
Best Regards,
MuhibUr Rahman
Author Response File: 
Author Response.docx
Reviewer 2 Report
The authors fuse features obtained from the input patch by different techniques (HOG, CN, etc.). Euclidean distance between these features, as used in Eq.8 may not have any meaning, at least the authors did not prove that it does. The reduction of features described further rely on the assumption that all the features are somehow balanced.
How the threshold used in Eq.10 was found? Empirically, or there is a reason to choose the value of 0.4?
With the present state of computational hardware does it make sense to reduce number of features by a factor of 4? What is the speedup achieved with this reduction?
The proposed technique outperforms other methods mostly by less than 2% (only in one case it is 5.3% - for success, where values are in 0.4-0.6 range). Could it be that without feature reduction it would perform better?
In section 4.3 the authors probably meant "ranges" rather than "rages".
In section 4.4, "summarizes" and "shows that".
Author Response
Dear Reviewer,
First of all, we would like to sincerely thank you for their thoughtful comments, questions, and suggestions. We hope that we satisfactorily respond to each one of them.
Attached files are as below
Revised manuscript.
Revised manuscript with marks for the updated sentences.
Reply to the Comment of the reviewer.
Thank you for your contributions to this manuscript,
Best Regards,
MuhibUr Rahman
Author Response File: Author Response.docx
Reviewer 3 Report
This article compares the precision of various types of multi-channel feature tracking models using different filter methods, and proposes a new descriptor for supposedly better precision and success rate in comparison to the state-of-the-art tracking models. This is achieved by reducing the computational complexity of existing approaches through a combination of Fast Fourier Transfom (FFT) with iterative feature learning. A large body of simulation data is compared within a visual tracking benchmark data set to show the improved accuracy of the new method in comparison with the other state-of-the-art tracking models/methods.
This paper contains potentially valuable work, but the way in which it is written poses several major problems:
1) The number of abbreviations in the article is overwhelming and, more importantly, quite a number of them are not explained and/or not even made explicit. For example, in the abstract the abbreviation SVM pops up without an explicit definition. The reader can only guess that the authors may (or not?) refer to Spatial Visual Mapping. The same problem occurs already in the abstract with respect to the terms OMTLBP and HoG. Then, in the introduction, the abbreviation PCA pops up in relation with dimensionality reduction and, again, the reader must guess that Principal Component Analysis is referred to here. There are more examples of this problem further-on in the manuscript. This kind of obscure writing and excessive use of abbreviations is not acceptable. The article needs to be revised page by page to make sure that 1) all abbreviations used are made explicit before they are introduced and 2) that abbreviations are avoided whenever this is possible. For example, the term 'colour naming' (and some other terms) should not be replaced by 'CN' everywhere in the text, but written in full throughout.
2) Apart from this problem of clarity in writing, there are major issues of a more general nature. First, the paper is submitted to the journal 'Symmetry' but does not address the implications of symmetry for differences in precision between the proposed method and the state-of-the-art tracking models for the benchmark dataset given. In other words: is symmetry relevant here? If yes, why and to which extent is it relevant, and how does the proposed method highlight such relevance further? Second, this paper is to go into a Special Issue on Deep Learning, but does not exploit deep learning algorithms, or fails to make clear that such are indeed exploited. As is, the paper states to exploit a Fast Fourier Transform (filtering) in combination with iterative feature learning, which is a traditional machine learning method. Deep learning involves, by definition, functional Artificial Intelligence architectures such as Convolutional Neural Networks (CNN), which are often also called Deep Neural Networks (DNN). These issues need to be resolved by major revision, or by submitting the article to a different journal if neither the role of symmetry nor the function of deep learning can be brought to the fore by revising the paper adequately.
Author Response
Dear Reviewer,
First of all, we would like to sincerely thank you for their thoughtful comments, questions, and suggestions. We hope that we satisfactorily respond to each one of them.
Attached files are as below
Revised manuscript.
Revised manuscript with marks for the updated sentences.
Reply to the Comment of the reviewer.
Thank you for your contributions to this manuscript,
Best Regards,
MuhibUr Rahman
Author Response File: Author Response.docx
Round 2
Reviewer 3 Report
This article has been revised substantially and the editors of the special issue may now decide whether it fits in. The paper is in principle acceptable for publication pending a minor spell and grammar check.
Author Response
Thank you for your contribution to this manuscript.
Comments and Suggestions for Authors
This article has been revised substantially and the editors of the special issue may now decide whether it fits in. The paper is in principle acceptable for publication pending a minor spell and grammar check.
Reply: We have double checked and improved the spelling and grammatical mistakes. We do hope that the revised version will be consistent and coherent.
