24小时热门版块排行榜

返回列表

61113377

新虫 (初入文坛)

应助: 0 (幼儿园)
金币: 38
帖子: 4
在线: 35.6小时
虫号: 2062764
注册: 2012-10-15
专业: 信号理论与信号处理

[求助] 第一次投SCI（PR Letters）,一个R，一个AQ，求助应该怎么办，已有6人参与

本人第一次投SCI，选择了PR letters，（当时听别人说的这个审稿时间比较快，就选了这个）结果是一个R，一个AQ，其中的很多意见都不知道怎么回答，
问题一：从审稿人的意见来看，改后再投还有戏吗？
问题二：请问如果大改后再投的话，是不是需要将文章的model和算法都要进行修改？
问题三：如果转投别的期刊，大家有什么好推荐的期刊（我是做CV中的行人再辨识的），影响因子低没有问题，主要是好中，然后审稿期不要太长（一年的那种我等不及啊）
我将审稿意见附在下面，请各位给我出出主意吧，谢了

---------------------------------------------------------------------------------------------------------------------------------------------------
** {This applies to SUBMITTING AUTHOR Accounts ONLY: You can find any possible ATTACHMENTS FROM THE REVIEWERS by going to the "Manuscript with Decisions" status link in your Author Center and clicking on "view decision letter". They are located at the bottom of decision letter under "Files attached" heading}

Dear Mr. Zhao,

The reviewing process of your paper submitted to the IEEE Signal Processing Letters is now completed.  Comments from the reviewers are attached at the end of this email. (** See note below about attachments).

Based on the attached set of reviews, I regret to inform you that I have to decide to REJECT the  paper for publication. Two highly qualified reviewers have looked at your manuscript in detail and, in general, they found it not acceptable for publication in its present form. While reviewers found some merit in the developments and experiments, they still find the paper requires a holistic major revision, and clarification of many issues. The first reviewer raises important problems on the approach, while the second reviewer asks for many non-trivial issues to be discussed in the paper. Considering that the decision process for the IEEE Signal Processing Letters is BINARY (papers that need major revisions are not accepted), I regret that I cannot offer you a more positive decision at this point because we do appreciate your interest in publishing in the IEEE Signal Processing Letters.

Resubmission of Previously Rejected Manuscripts:  Technically, you cannot resubmit a REJECTED manuscript, as it is a REJECTED and CLOSED paper.  You would therefore need to submit it as a new manuscript obtaining a new manuscript ID #, following the guidelines in the Author Center (where you would submit your paper to the system) under "RE-SUBMISSION OF A REJECTED MANUSCRIPT"

Authors of Rejected manuscripts are allowed to resubmit their manuscripts only once. The Signal Processing Society strongly discourages resubmission of rejected manuscripts more than once. At the time of submission, you will be asked whether you consider your manuscript to be a new submission or a resubmission of an earlier Rejected manuscript. If you choose to submit a new version of your manuscript, you will be asked to submit supporting documents detailing how your new version addresses all of the reviewers' comments. Full details of the resubmission process can be found in the Signal Processing Society “Policy and Procedures Manual” at http://www.signalprocessingsocie ... e/policy-procedure/

Note that resubmitting your manuscript does not guarantee eventual acceptance, and that your resubmission will be subject to re-review by the reviewers before a decision is rendered.  Also note that the original Associate Editor who managed the original peer review process is not guaranteed as well.  Resubmissions are to be treated as brand new submissions without bias.

Sincerely,
Prof. Gustau Camps-Valls
Associate Editor
gcamps@uv.es, gustavo.camps@uv.es

* If you have any questions regarding the reviews, please contact the managing Associate Editor who managed the peer review of your paper.

Reviewer Comments:

Reviewer: 1

Recommendation: R - Reject (Paper Is Not Of Sufficient Quality Or Novelty To Be Published In This Transactions)

Comments:
The paper presents a method that looks for addressing the problem of matching images that have similar features. They are focused on match images of pedestrians which have been taken with two different cameras. The method is based on learning a specific metric for each datum.
In general I see the proposed solution too much complicated, one have to fit a huge amount of parameters. Besides the general parameters: mu, beta, c and P. For each pair of data, one have to find, Wx-> dxK, Wy-> dxK, L-> dxKd. I think that this amount of parameters make the method suffers in overfitting. Although I have my concerns about the method, I would be able to recheck it, if the authors remedied all the issues below.

MAJOR ISSUES
- Regard the test procedure.
Authors say:
"At the test time, for a given test image, we find its P-nearest images in the training set and assign the weighting vector of the test image as the linear combination of these P corresponding weighting vectors."

This is the critical point of the method. Is it robust for new test images? At least, it does not seem so. For instance, if a test image has not similar images in the training set, the linear combination of the weights of the assigned training images would have no sense. Authors has to justify the robustness of the method for new samples.

- Regard the parameters employed.
Authors say:
"The parameters in our method are set as μ = 0.5, K = 10, β = 2.5, c = 1, P = 5."

Authors have to explain which method they have employed to select these values.

- It is necessary to present the gradient function for obtaining the solution of the W matrix.

- The code of the method has to be publicly available.

- Authors have to justify the sentence:
"In order to accelerate the optimization process, we design an initial value of weighting matrix as follows."

MINOR ISSUES

- LCC is not defined in the abstract.
- Please change. Recently, a multi-metric based method LAFT [14] is ...
-> Recently, a multi-metric based method LAFT [14] was ...
- Please change. ...work are presented in Section IV.
-> ...work is presented in Section IV.
- Page 2. Define \mu before (or just after) equation 5.

Additional Questions:
1. Is the topic appropriate for publication in this transaction?: Yes

2. Is the topic important to colleagues working in the field?: Yes

Explain:

3. How would you rate the technical novelty of the paper?: Somewhat Novel

4. How would you rate the English usage? : Satisfactory

6. Rate the references: Satisfactory

null:

Reviewer: 2

Recommendation: AQ - Publish In Minor, Required Changes

Comments:
Summary: This paper presents a locally adaptive method for person re-identification. Conventional supervised methods for re-id train a single verification model, transform or mahalanobis metric to improve matching accuracy across the views. However, in reality the transform between the two views is likely to be multi-modal, so a single metric is not ideal. This paper builds on the ideas of local coordinate coding and metric learning to build a metric that is adapted to any given pair of points to be matched. The results are competitive with state of the art methods on three benchmark datasets.

Overall:
+ The notion of customising the matching method according to the specific pairs to be matched is powerful but currently under explored in re-id. The main existing approaches are [14], and (un-cited) [A], [B] However this seems to be quite different to both, so is a worthwhile contribution.
+ The performance is comparable with current state of the art.

However, a few issues should be fixed before acceptance:

- The dismissal of mixture of experts due to "partitioning the space in a complex way" is a bit too cursory. (Especially since [14] still beats the current work in some settings).
- It seems that this paper has 2NxK + dxKd parameters to learn for the weights and the projections respectively, compared to dxd for normal mahalanobis metric learners. This seems like a lot of parameters. Please give some intuition about why over fitting isn't a problem. Especially since MoE is criticised as risk of over fitting.
- "modalities" seems to be used to essentially mean camera views. This is confusing, as modality more typically means sensing modalities such as color/texture, image/vision.
- P1L59. Is [19] anything to do with multi-metric? Is this the wrong reference?
- Around Eq (4), other subscripts are indexes. But I think 'b' in L_{bk} just means basis projection, rather than an index, so maybe use superscript or something else to indicate a basis projection.
- Eq(10), shouldn't it be d(x,y) on LHS, not d(\bar{x},\bar{y}) ?
- The test time procedure is insufficiently clearly explained. For ease and logic of reading it should be a separate Sec III.C, rather than buried in a paragraph of III.A. Then please explain it in a bit more detail. Why is the test image encoded as NNs in training? (Why not also encoded also in terms of the LCC?) And how are the weights in the linear combination determined? And what is the intuition about the significance of the choice of P?
- Citations [A],[B] should be included and discussed/contrasted.
- The original LCC paper seemed to learn the anchor points and the weights jointly. Here the anchor is pre-computed by KNN, and then only weights are learned. Some discussion is necessary about what is lost by doing this in terms of the favourable properties/guarantees of LCC.
- Sec IV says the feature vector is d=5895 dimensions and K=10. It seems improbable that a Lbar matrix of size 5895 x 58950 can be learned. I suppose there is some missing dimensionality reduction step? Please explain or fill in the missing details.

Please check english thoroughly, there are many errors. To mention just a few:
- Abstract P1L19-20 makes no sense. Potential to partition?
- P1L25 liner -> linear?
- P1L37 attention(s)
- Various places: "generated by clustering method" is wrong. By a clustering method, or by clustering.
- P4L39-40 non-standard english.
- P4L20 sufficiently characterise(*) the data
- P4L21 even if they come.

[A] CVPR'13: Learning Locally-Adaptive Decision Functions for Person Verification.
[B] Pattern Recognition 2014: On-the-fly feature importance mining for person re-identification

Additional Questions:
1. Is the topic appropriate for publication in this transaction?: Yes

2. Is the topic important to colleagues working in the field?: Yes

Explain:

3. How would you rate the technical novelty of the paper?: Somewhat Novel

4. How would you rate the English usage? : Needs improvement

6. Rate the references: Satisfactory

回复此楼