Weifeng Chen@Zhejiang University<chenweifeng@cad.zju.edu.cn>

2010-09-14 15:06:13

Thank you for the comments on our paper. But we totally disagree that our work is a copy of "Coordinates for Instant Image Cloning" paper, which uses Mean-Value Coordinates (MVCs) for image cloning.

First, our method solves one main limitation of the MVCs work at cloning concave regions. Please refer to examples in Fig. 5. We compare our method with the MVC and visualize the difference. It can be observed that our method produces more smooth and natural cloning results. It is because, in mathematics, our Harmonic coordinates has better capability for concave cases that the MVCs is less successful. We only provide one comparison is because we think one image is able to clarify the difference of two methods. If it requires, we’d like to put more comparison on the web. The analysis of advantages of our HCs than MVCs is presented in section 3.2. The author of the MVC based cloning paper argues that “cloning with highly concave regions (when MVC performs badly) is hardly ever needed”. But, in an interactive image cloning system, nobody is able to expect what is the next shape or region user wants to clone. It is worthy to exploring a new method to overcome the shortage of the MVCs based cloning and provide better solution.

We also noticed another image cloning paper (“Improved coordinate-based image and video cloning algorithm”, Lee, S. and Lee, I., ACM SIGGRAPH ASIA 2009 Posters) aiming at solving the same problem. Different from our method that employs a new coordinate, it uses a new scheme to compute the coordinate weights. Both of our works make steps on improving the MVCs. Although we all agree the "Coordinates for Instant Image Cloning" is a great paper, there is no reason to stop the paces to explore better coordinates and solutions for image cloning.

Second, our main contribution is the Harmonic coordinates for image cloning. We provide our solution to compute the coordinates and present a GPU-based Harmonic coordinates solver. In section 4 we propose the equations to generate the coordinates and in section 5 we give the GPU-based implementation to solve the Poisson equation. Based on these contributions, we are able to achieve better quality of results and still retain real-time or interactive image cloning. As an extension and improvement of MVCs, our method requires including some algorithms in MVCs to make the representation of our paper clear and complete. But, to clarify the originality and novelty hence avoid the confusion and misleading, in our paper, we explicitly cited the MVCs paper and clearly claimed in necessary places which parts are from the MVCs paper and which parts are our adaption. We do not regard or intend to regard algorithms proposed in MVCs paper as our contributions. The adaptive mesh, hierarchical boundary sampling and the temporal smoothing for video cloning are all marked out from the MVCs paper. It is unfair and unjust to make censures but ignoring our contributions and explicit citations.

Third, we presented our work as well as the MVCs paper in the same session on “The 1st China-Israel Bi-National Conference on Graphics and Geometric Computing”, 2009. (Please refer to http://cg.cs.tsinghua.edu.cn/ci2009/?page_id=4 for the conference schedule.) We exchanged comments and opinions with many academic peers including one author of the MVCs paper in the conference. However, we did not receive any comments on regarding our work as a copy of the MVCs paper in the conference or after it. It is surprising for us to read such a comment here. We hope our work can be evaluated justly.

In summary, our paper provides an alternative coordinates method for image cloning. Although our work is based on the MVCs paper, our new coordinates and fast GPU solver enables a better image cloning especially for cloning concave shapes or regions.

Gil Hoffer@Tel Aviv Uinversity<gilhoffe@cs.tau.ac.il>

2010-09-11 16:01:33

This paper is almost an exact copy of our "Coordinates for Instant Image Cloning" (Farbman et. al.) paper.
There are complete paragraphs which are almost identical, and many many algorithmic details which are exactly the same (using an adaptive and hierarchic scheme, the temporal smoothing for video cloning, etc...)

Also, you don't really give a good example for why your HC based system performs better than easy-to-implement MVC based cloning. From our experiments, cloning with highly concave regions (when MVC performs bad) is hardly ever needed (for instance in your example in figure 5, the more sensible choice would be to use a convex disc-like shape for the concealment)