Thanks for your encouragement. I believe that I can manage it. Nothing can stop my steps towards the future. From now on, work hard, play hard :-)

 

 

 

今天的TOEFL考完了。听力做得很郁闷，好些题目听得似是而非，分数很低。

 

因为是机考，当时就给出一个成绩范围，看了就很郁闷了。因为作文是折算进总分的，要达到标准，作文起码要5分。虽然，我相信我作文能力不止5分的水平，但是，考试的事情真的不好说了。起码，在非作文部分，我取得的分数比预想水平低了十多分。

 

希望作文能拿到5分或者更多。这样，这一关算是勉强过了。

疲乏，这是这几周来最主要的感觉了。几年前，这还是一个遥远的梦想，今天，它已经是一个实实在在的目标。一路上的奋斗给了我太多的回忆，即将到来的TOEFL机考，将是这条路上最后一个关口了。曾经的满怀憧憬，在彼岸似乎触手可及之时，却变成了不安和焦虑。似乎胸有成竹，现实却依旧千头万绪。

 

备考其实不是我在香港最紧张的日子，可是比起最紧张的conference deadline前的彻夜不眠，更容易让人疲乏。毕竟，paper是一件创造性的工作，它能让你满怀热情地投入战斗；可是，托福的最后阶段就是反反复复练习那些类似的东西，无休止的重复性练习。对我来说，最希冀的不是在最后的现场考试拿到多高的成绩，而仅仅是停止这种也许可以叫做折磨的生活。

 

下个月中下旬，有好多人都要到美国了，青青要去西雅图，小呆去匹兹堡，老板，还有我要去纽约。虽然这么多朋友在那里，遗憾的是，我们很可能不能聚首。在异国他乡的孤独，不是一件让我向往的事情。

Recently, I am in retrospect to a question that what on earth is a good paper. In the past, it is my tendency to develop sophisticated systems with multiple advanced components integrated. Though not explicitly admitted, I prefer the papers filled with complex formulas using advanced mathematical symbols. However, the conferences held recently motivate me to rethink the true value of an academic paper.

 

Generally speaking, the ultimate aim of a paper to address a problem that has not been addressed. In this sense, a good paper should be dominated by its main goal, but not what advanced components it has. For a newcomer of research, his enthusiasm is prone to be devoted to some fantastic models or methodologies, and his thoughts would be accordingly restricted around. I was deeply influenced by this notion. You can see some of my papers, even those that are published in high-level conferences such as CVPR and ICCV, pile up various fashionable techniques. However, I gradually realized that this is not a correct way. As pointed out by some reviewers, these approaches lack elegancy.

 

Seen from the history of the evolution of computer vision, the great works that truly have a durable life and broad impact are those that bear a simple and elegant core, such as PCA and SVM. The flowery and complicated designs may add some superficial values to a paper. However, in a long term perspective, the true significance originates from the simple core that makes it work.

 

In my view, a successful paper relies on the following keys:

 

1. A novel core. Prof. Tang have said for many times that the contribution of a good work should be able to summarized by a short sentence, which is what I refer to the core of a paper. It is impossible for a paper to resolve all the problems in a domain, so it is typical to use some existing techniques in the paper. No matter how many new and advanced techniques you use in your paper, they contribute little to the novelty of your work. We should keep in our mind that what endows your paper with true value is that has not been done.

 

2. Clear and focused target. Each paper should focus on a single and well-defined problem. All components in the framework serve it. I don't think it is a good way to add things irrelevant to the heart of the paper in order to "enrich the paper". In the eyes of an experienced reviewer, this not only offers no extra values, but also distracts the readers from the core and essentially impairs the significance of the paper.

 

3. Crystal clear presentation. Don't be proud of a paper that can not be understood by most people. The object that you publish your paper is to make your work applied in practice or exchange your thoughts with other researchers, but not to flaunt your mastery of advanced theories. When I review papers, I will not hesitate to reject any papers that are difficult to follow. I believe that many other reviewers share the same idea with me, after all no one would like anything make him confused.

 

4. Easy to implement. It is widely acknowledged that an elegant solution generally involve an efficient and simple procedure. A case in point is a classic algorithm called PCA in statistical learning. Though it is old, it has a great impact in this field, which is to a large degree due to its simplicity and efficiency in implementation. The procedure with many complicated steps can by no means be taken as a beautiful design. It is my experience in getting my papers reviewed and attending conferences that most people show their favor to simplicity.

Maybe I don't have enough time to write another summary. So I just posted the mail I wrote to the lab here.

-------------------------

 

I have come back from the 9^th ECCV conference. It is really a rewarding trip in Graz.

 

It is my opinion that the two presentations given there is basically successful; some researchers have exhibit interests in our works. Thanks for the lab on giving me such a valuable opportunity to communicate with the researchers all over the world. I would like to give a brief summary to the conference here.

 

During the conference, I feel that the most heated topic is some so-called “true sense” vision problem, such as segmentation, stereo and tracking. It is reported by the program chair in the banquet that the papers in these topics have relative higher acceptance rate and more presentations are given on these topics. The best paper award has been granted to a paper on segmentation:

 

Learning to Combine Bottom-Up and Top-Down Segmentation

Anat Levin, Yair Weiss.

 

In Anat’s presentation, she states that the major contribution of the work is to combine the top-down models and bottom-up cues in the learning stage, instead of testing stage as in some previous works. The learning of the representation of low-level fragments and high-level models are done simultaneously with information exchange.

 

In addition to the best paper, two papers won the honorable paper awards:

 

Confocal Stereo

Samuel W. Hasinoff, Kiriakos N. Kutulakos

 

Simultanesou Object Pose and Velocity Computation Using a Single View from a Rolling Shutter Camera

Omar Ait-Aider, Nicolas Andreff, Jean Marc Lavest, Phillippe Martinet

 

Though my major interest is not in these topics, I have listened to a series of presentations in these papers. It is a conspicuous trend in these works, the combination of low-level cues and high-level models. Though they are formulated in various forms, the thoughts of such a combination dominate the trend. Please note that in many good works, the combination is not a simple framework to use A method in first stage and B method in second stage, but a coupled formulation given in a creative manner.  I think it is helpful for us to learn something from these works about how to develop interesting and interdisciplinary works by connecting different topics inherently.

 

For recognition, including face recognition and object recognition, I think the topic on local features receives surprisingly much attention in the conference. Many new local feature extraction techniques have been proposed. Learning the hierarchical structure are introduced into local feature extraction and representation. The statistical models based on local features become very hot. However, the enthusiasm in learning seems cool down. I don’t see very exciting works in learning in this conference. That is not to say learning becomes less important. What I mean is that the development of new learning methods is not very hot in the conference; nonetheless, various models using learning are more widely applied in most of the works.

 

In sum, the research on some traditional topics become more and more difficult, while the efforts in opening new topics by bridging different domains would gradually become a considerable force to revive the realm of computer vision. I think it is high time we broadened our scope and absorbed new notions from other fields instead of restricting our views in a small domain.

 

In addition to the awarded papers, several oral papers worth recommendation:

 

Controlling Sparseness in Non-Negative Tensor Factorization.

Matthias Heiler, Christoph Schnorr

 

NTF (Non-Negative Tensor Factorization) is a new technique in learning. It is a tensor-based generalization in NMF (Nonnegative Matrix Factorization). The authors present a solid work in developing a optimization framework in tensor factorization with sparseness constraint. The algorithm may produce global optima in the cases satisfying some conditions. This work may be enlightening in both sparseness constraints, convex optimization formulation and tensor factorization.

 

Figure/Ground Assignment in Natural Images

Xiaofeng Ren, Charless Fowlkes, Jitendra Malik

 

It is a very good work accomplished by Ren Xiaofeng, UC Berkeley. At least, his presentation is the best one I have ever seen. The work is on a rather new topic: foreground and background assignment. Different from the topic on segmentation, it does not aim at segmenting the images into regions, instead it is to tell which region is foreground and which region is background given a segmented image. They accomplished this task based on the shape hints on region boundary instead on the contents within the region. Briefly, the shape primitives and their relation to the foreground/background judgment is learned, and constitute a probabilistic network.

 

Level-Set Curve Particles

Tingting Jiang, Carlo Tomasi

 

It is a good work in connecting the level-set and particle filtering.

 

 

Just now I have completed my second oral presentation in ECCV 2006. It is basically satisfactory, the presentation is given very fluently.  In the album there are photos  taken for me during my presentation.

Graz is a beautiful city in Austria. You will not see many modern buildings here, most of the buildings date back to many years ago, even to the era of renaissance. Actually, walking along the streets here brings me back to some hundred years ago. The photos in the MSN album will show you the street around the congress venue. They are not in a historic park, but in a typical street here. Through these photos, I hope to share with you my wonderful experience in Graz.

 

However, in my opinion, it is a place only suitable for short term trip, but not for long-term life for me. First of all, I am not get used to the foods here. It is indeed a big problem when I intend to find something to eat. They are not delicious and very expensive. Moreover, many people here know little English, the communication with them is not very convenient. The text on the boards and posts here is not in English but in German. I can hardly find an English word.

 

Well, I think you cannot wait to read more. Now please go to the album for a virtual journey in Graz.

It is the first day of the main conference of the 9th European Conference on Computer Vision. In the morning, I have given an oral presentation for one of my paper on Conditional Infomax Learning. Though I did not have sufficient time for preparation, it is my feeling that the presentation is given fluently. However, during the lunch time I have heard some not very good comments on my presentation. A guy said that the presentation is not logically organized, and is difficult to follow. Maybe he is right, I have thought that the time for presentation is only 15 minutes, which is obviously insufficient to cover all major aspects clearly. However, before I was giving the talk, I wad informed that we can have 20 minutes. In the following two days, I need to re-organize the oral presentation for my second paper to cater for the new requirement.

Oh, the reception time comes. I need to leave.

几个小时后，我就要登上飞机，去往一个从没踏足的国度——奥地利。在我的印象中，奥地利仅仅是地理教科书上给我留下的概念，一个欧洲国家，也许还有一点历史书上的痕迹——哈布斯堡王朝和它治下的德意志神圣罗马帝国。而它的首都，维也纳，在我的感觉中是一个遥远而美妙的地方，近代欧洲的那些伟大的音乐家们，海顿、贝多芬、莫扎特和施特劳斯，给这个本来就很典雅的名字赋予了浓郁的音乐色彩，让它成为一个让无数人心驰神往的艺术之都。

会议是在奥地利一个古老的城市召开的，叫做格拉茨——我在此之前没有听说过的城市。最近，看了点关于它的文章。这个城市充满了欧洲各个时代风格的建筑，从中世纪的城堡，文艺复兴时期的教堂到近代的。 其中的格拉茨艺术馆更是久负盛名。由于这个城市的古建筑保存完好，1998年，格拉茨获选“欧洲文化之都”；1999年，格拉茨城历史中心登录世界文化遗产名录。有趣的是，曾经的好莱坞著名演员，现在的加州州长施瓦辛格也出生于此。

这次旅程最核心的任务自然还是要去present自己的论文。汤老师说，在这种世界级会议上两次登上oral的讲坛，是难得的荣誉。在等待出发的期间，内心如一镜湖水的平静，可是一种向往，一种豪情，依然不时在湖水中激起涟漪。

这次旅程虽然很短暂，只有不到10天，还是在这里向大家道别，再见了。

申请万事俱备，只欠TOEFL了。要赶上旧TOEFL的末班车。

 

 

2006年5月1日，下午。

 

今天是国际劳动节，在国内，这是一个人潮涌动的黄金周首天，香港也放假了。诺大的校园空荡荡的，只有烈日的炙烤。

 

对于我来说，自从来了香港，就没有所谓节假日的概念了。当朋友们问我五一放不放假的时候，我告诉他们，所有的日子都是假日，所有的日子都不是假日。在我们实验室，每个人对时间支配的自由度之高，无论是在国内还是国外的同学都感到惊讶。假日对于我来说，最重要的不同，就是范克廉楼的餐厅几乎都停止营业了，我需要付出更多的脚力才能填饱肚子。

 

今天是断不敢进出深圳的。报道说，今天早上开关后不到5个小时，就有5万人通关进入香港，全天会有20万人人流。我在春节和复活节就亲眼目睹过成千上万的人堆满整个边检外广场等待进关的壮观情景。即使在香港，节日逛街也不是一件好差事。在节日到过旺角，就知道从人缝中穿插过好多条街的感受了，摩肩接踵似乎已经不足以形容这种拥挤.

 

几天后就要到欧洲去了，因此这个假日还要忙完很多手头的事情，千头万绪。要准备测试平台，准备买数据采集设备，准备去ECCV的presentation，准备考TOEFL，准备写paper，准备毕业论文和答辩， 准备办各种很累又不得不办的手续，恩，还要写完这片blog，还要每个晚上骚扰朋友们和被朋友们骚扰——没有了这种骚扰，生活也真是太寂寞了。

 

假期，就在繁忙和寂静中度过。不过，充实总比空虚要好。