Posts from February, 2009
No comment yet
February 26th, 2009

曾经是魏晋的遗风,国人直到现在仍然笃信语言的力量。和当时的清谈一样,大家相信言语胜于行为。比如如果一个人说自己是老实的,那么潜在地,大家会认为他一点都不老实,甚至当他的行为表现符合老实的定义时,还有很多人根据这样一句话认为他不老实。懂得运用语言来混淆视听,即使在当代,也有着巨大的作用。因言既然能够获罪,那么言语也能够惑主,也能够获利了。

我并不是在吹捧孔子所说的言伪而辩的概括是多么精辟。毕竟,不能否认语言在交流中所起的作用,通过辩论来证明事物也并不错误。问题在于人们广泛地将语言和实际事物相联系,却没有意识到语言知识媒介,和实际的事物没有任何的关系。

这就是为什么三人能够成虎的原因。本来有虎这样的一句话是没有任何意义的,他不代表有虎或者没有虎。而人们试图通过个人信用度为其担保,最后在没有任何印记的情况下虚构了一只老虎出来。这是所有原始传播媒介的悲哀,由于本身不带任何担保,使得信息可信度不能科学度量。

听其言,观其行,这在任何时候都是真理。

No comment yet
February 14th, 2009

社会发展的平衡和不被特殊事件影响的可能性取决于个体的发展。简而言之,个体发展程度越高,文明将会越不稳定,最终会由于少数的事件导致整个文明群体的灭绝。

这一论断的得出基于这几点观察。一、随着文明发展,个体能获取的能力将会越来越大。二、在整个群体中,个体的活动是近似随机的。从第二点出发,我们知道个体的活动函数f分布有随机性保证,但是群体活动F将是个体运动f的和,是稳定的。设定每个个体活动函数f受到他的邻近个体活动函数f{k}的影响,且||f-f{k}||小于一个阈值c。由于第一点的观察,阈值c不断放大,最后导致每个f均受到整个群体的显著影响。由于每次计算受到时间间隙的影响,即使最终的F是稳定的,在每个时间间隙点仍然是不稳定的,若假设个体活动函数f小于某阈值z时表示死亡且不可逆转,那么在每个f均能影响整个群体时,群体灭绝性事件的发生是不可避免的。

这一论述为未来文明的发展说明了几种可能性。利用行星系尺度的间隔,可以减小个体影响力的作用,即使一部分世界会由于少数事件而灭绝,从概率上说,少部分世界仍然可以存活发展下去。减少个体数量,在这种情况下,群体发展会更加不稳定,但是个体行为更加容易预测,一个随机模型将不再适用,整个预测模型的失效可能会产生新的可能性。一种减少个体数量的极端办法是使得整个群体具有能力,而不是个体。实用地说,通过制度的设定保证个体无法产生对群体具有显著影响的能力。

现阶段的危险性在于,通过各种辅助方式,个体的能力已经显著提高。以前,操作船舶需要几十个人的配合。而现在只需要一个人的控制和计算机辅助系统就能驾驶飞机。个体具有对整个文明产生决定性影响的能力在将来的不久就会出现。非对称化的恐怖袭击只是这一论断的一种表现而已。

No comment yet
February 5th, 2009

搞MSER搞了一周了,还是在找local minimum的时候有问题,而且还比原文的实现结果慢了4倍左右,感慨怎么实现点稍微复杂的算法就已经这么疲惫了。

No comment yet
February 4th, 2009

Local feature descriptor (LFD) is an overwhelming successful method for image comparision which is currently the best solution against in-plane-rotation, distortion and light variations. However, there are some assumptions should be noticed. LFD is a appearance feature. It is more stable than pixel feature, but after all, the property of appearance feature still disturbs the ability of LFD. First, the appearance feature is vulnerable to variations of light and the description ability is depended by the complexity of appearance. LFD by collecting features through the key areas counteract the interferences of image variations. You can view that as sort of extension to shape description. Still, for object with low complexity of appearance, LFD failed to achieve any thing. The test senerio will be a color ball with a complex background. For this senerio, LFD cannot capture any useful information of the ball because the distinguish of the ball is the shape not appearance. For that part, LFD can do little.

One way to solve the problem is to combine some sort of shape descriptors to our LFD machanism. As appearance features and shape features are total different in their domain, there is no simple way to do this. There are rare cases that shape is more powerful than appearance. A closer look at how shape affects the appearance we will notice that only smooth boundary is something LFD cannot describe well. For sharp boundary and turnning point (corner point), LFD algorithm can effeciently extract local descriptor from those key point. In our extreme case, LFD lost its most important function to keep track the percise location of key point. For a ball, there is just no way to judge which point is key point on the boundary of circle.

Another limitation that is heavily addressed is the computing complexity of LFD. In my previous articles, I provided some insights of how to solve the problem.