--------------------------------------- TIP 2013 (accepted) --------------------------------------- ------------------- Revision #1 ------------------- --------- Area_Editor (Jenq-Neng Hwang) --------- We are pleased to inform you that the above paper has been ACCEPTED as a regular paper in the Transactions on Image Processing. --------- Assigned_Reviewer_1 --------- Recommendation: A - Publish Unaltered Comments: In the modified version of this manuscript, the authors have given detailed and reliable answer to each question raised by reviewers. Many changes have been made to increase the readability of this article. Most importantly, a lot of experimental results have been added based on several widely known and used datasets, allowing for a direct comparison with other methods. The detailed analysis of the strengths and weaknesses of the adopted solution was added. --------- Assigned_Reviewer_2 --------- Recommendation: A - Publish Unaltered ------------------- Original submission ------------------- --------- Area_Editor (Jenq-Neng Hwang) --------- Based on the enclosed set of reviews, I am recommending that the manuscript be REVISED AND RESUBMITTED (RQ). (Major revision) --------- Assigned_Reviewer_1 --------- Recommendation: RQ - Review Again After Major Changes Comments: This paper introduces a method for logo detection based on local features and their geometric context. Using the definition of context and similarity design from [39][40], the proposed method seeks to obtain accurate image matching results by minimizing an energy function, which is claimed to have good tolerance to partial occlusion of logos in different images. In addition, the authors also provide the theoretical foundation of the matching framework which shows that under the hypothesis of existence of a reference logo into a test image, the probability of success of matching and detection is high. However, this paper has several significant shortcomings: (1) The novelty of this paper is limited. Several spatial/geometric context description methods have already been proposed in many different papers. They all use similar principles to define spatial context, such as the mentioned references [27]~[36]. As for the feature set matching strategies, there are also many existing optimization algorithms. Proposed method is only compared to several strategies presented around 2003, not to state-of-the-art logo detection methods like [27]~[36]. (2) This paper is an incremental improvement over [39][40], and at least 60% content are almost the same. The main difference lies in that context is used for logo detection in this paper, while context was used for object classification in [39][40]. However, logo detection task mentioned in this paper already have many efficient and effective solutions with high detection precision. Different with classification applications, the time-consuming energy minimization calculation might not be needed in logo detection. The same or better performance may be achieved by systems that use faster and less-complex matching. It's not clear whether the proposed approach should be preferred in practice to state-of-the-art logo detection systems like [27]~[36]. (3) Most importantly, the experiments are performed on a non-standard logo detection dataset. In practice, there are often thousands of logo categories in a dataset, and computation of energy minimization may be very slow and laborious for logo recognition task. But the dataset used in this paper only contains 13 logo types and 720 relative images in all, which is difficult to demonstrate convincingly that the proposed method is practical in logo detection applications. In a word, the authors haven't proven that their system improves over the-state-of-the-art logo detection strategies, which is by now a very mature and well-studied area. --------- Assigned_Reviewer_2 --------- Recommendation: RQ - Review Again After Major Changes Comments: Please see the review on the attached PDF http://www.lambertoballan.net/downloads/2013_TIP_reviewer2.pdf --------- Assigned_Reviewer_3 --------- Recommendation: A - Publish Unaltered Comments: This paper provide enough introduction and literature review. Especially, it describes plenty of math proof for simiarlity mesaure. In experiments, the proposed method shows robust. The format of references should be improved.