• What Makes a Good Teacher
                What Makes a Good Teacher? 2004年05月26日10:42:39　解放日报--上海学生英文报　  　　这是一群美国三年级学生发表的观点。充满童趣。　　　A teacher helps kids learn to write small(写小字). A teacher takes us outside to play. Don't talk back to the teacher. Teachers do not like pink slips. Teachers will get mad if somebody is being bad. A teacher will always love her kids. Teachers are very very cool. 　 　—p.r. 　　A good teacher needs to sleep. A good teacher needs to read books so they know what to teach. A teacher has to be healthy. The student has to follow the rules so the teacher won't get in trouble. 　　 —Mac Pon 　　A good teacher is someone who listens to someone who cares and gets apples, peaches and pears. A good teacher is someone who's never late for school. She's nice to her kids and really is cool. Most of all a good teacher loves her students. 　　 —Meter Bogschanir 　　A good teacher gives homework and math to us. She lets us have recess (课间休息). A good teacher teaches us how to write. A good teacher reads to us. A good teacher makes us smart(聪明), and helps us go onto fourth grade. 　　 —Yela Gan 　　A good teacher teaches good things. A good teacher is nice and helpful. She gives us parties and gives us stuff(东西) to learn. 　　 —Burne Agations 　　A good teacher is organized(井井有条). A good teacher cares and is always ready for his or her students. A good teacher is always fair(公正)and respects(尊重) their students. A good teacher always works. 　　 —Felar Lader 　　She can teach you good things. She is kind and nice. 　　 —Thomas Pether 　　A good teacher makes a good listener. And a good teacher is ready for their students. A good teacher is someone that helps you. A good teacher is a person who lets us eat. A good teacher is one who lets us read and reads with us. 　 　—Barna Twist 　　A good teacher is nice. A good teacher is organized and gives a lot of work. A good teacher gives homework and makes you think hard. A good teacher makes you read a lot. 　 　—Kip Gella 　　They let us read books. They do not hit us. They let you eat in the classroom. They let you talk. 　　 —Pat Smith 　　A good teacher is one that gives you homework. A good teacher makes sure that you listen and will help you learn. A good teacher gives you free time to play. A good teacher makes sure that you're on time. A good teacher makes sure that you have everything you need. 　　 —Shane Lan 　　They let you go outside and give you a lot of work. They teach you stuff and do not hit you. They say nice things to you. They give lots of parties. 　 　—s. 　　A good teacher makes everything fair. They do not give to much work. They are fun and give some hard and easy work. They are not mean(坏，狠). 　 　—May Gell 　　A good teacher is firm. They have good manners and they don't do bad things. They buy things for you and they buy you toys, books, and things. They do not say bad words. Some teachers give you cake and ice cream. They have good ideas. 　 　—Jial Teel 　　What is a good teacher? I think a teacher should be good at explaining (解释) things. A teacher should be fun (有趣) but challenging(有挑战性). A good teacher should be organized(井井有条). I like a teacher who teaches fun activities(有趣的活动) in class and not just writing on paper. A teacher should like teaching. 　　 —Sela Cont 　　A good teacher teaches so many things. They help me to read a book. A teacher always teaches good things to the students. 　 　—Mario Back              再分享一下我老师大神的人工智能教程吧。零基础！通俗易懂！风趣幽默！还带黄段子！希望你也加入到我们人工智能的队伍中来！https://blog.csdn.net/jiangjunshow


展开全文
• What makes a good conversation.pdf
• What-makes-it-page资源分享
• Papers J. Hosang, R. Benenson, P. Dollár, and B.... What makes for effective detection proposals? arXiv:1502.05082, 2015. arXiv @ARTICLE{Hosang2015arXiv, author = {J. Hosang and R. Bene

Papers

J. Hosang, R. Benenson, P. Dollár, and B. Schiele. What makes for effective detection proposals? arXiv:1502.05082, 2015.arXiv@ARTICLE{Hosang2015arXiv,
author = {J. Hosang and R. Benenson and P. Doll\'ar and B. Schiele},
title = {What makes for effective detection proposals?},
journal = {arXiv:1502.05082},
year = {2015}
}

项目地址：https://www.mpi-inf.mpg.de/departments/computer-vision-and-multimodal-computing/research/object-recognition-and-scene-understanding/how-good-are-detection-proposals-really/

大纲

根据文章的描述顺序，以下内容大概会是：
回顾object proposal（以下简称为OP）的各种方法，将其分类。讨论不同OP在图片被扰动之后的在复现上的鲁棒性讨论不同OP在PASCAL和ImageNet上的Recall，这里作者提出了Average Recall（简称AR）的一种新的标准讨论不同OP对于实际分类的性能比较（用了DPM和RCNN这两个著名detector进行比较），以及说明了AR是一个跟性能相当相关的标准。
先上一个效果的一览表格：
注意到这里只列出了可以找到源码的方法，那么，下面一点点开始整理。

各种OP方法的回顾

作者大致将OP方法分成了两类，一类叫grouping method，一类叫window scoring method。前者是指先将图片打碎，然后再聚合的一种方法，比如selective search。

后者是生成大量window并打分，然后过滤掉低分的一种方法，比如objectness。另外还有一些介乎两者之间的方法，比如multibox。

Grouping proposal methods

作者将grouping的方法继续细分为三个小类。SP（superpixels ），对superpixel进行聚合；GC（graph cut），使用种子点然后group
cut进行分割；EC（edge contours），从边缘图提取proposal。下面分别调一个进行介绍

• SelectiveSearch (SP) [15], [29]：通过贪婪地合并超像素来产生 proposals。这个方法没有学习的参数，合并超像素的特征和相似函数是手动设定的。它被 R-CNN 和 Fast R-CNN detectors [8], [16] 等最新的目标检测方法选用。无需学习，首先将图片打散为superpixel，然后根据人为定义的距离进行聚合

• RandomizedPrim’s (SP) [26]：使用类似与SelectiveSearch 的特征，但是使用了一个随机的超像素合并过程来学习所有的可能（probabilities）。此外，速度有了极大地提升。

• Rantalankila (SP) [27]：使用类似与SelectiveSearch 的策略，但使用了不同的特征。在后续阶段，产生的区域用作求解图切割的种子点（seeds ）（类似于CPMC）。

• Chang (SP) [38]：结合 saliency 和 Objectness 在一个图模型中来合并超像素实现前景/背景（figure/background）分割。

• CPMC (GC) [13],[19]：避免初始的分割，使用几个不同的种子点（seeds ）和位元（unaries ）对像素直接进行图切割。生成的区域使用一个大的特征池来排序。随机初始化种子点，然后做graphcut进行分割，反复多次，然后定义了某个很长的特征进行排序。（所以速度超级慢）

• Endres (GC) [14], [21]：从遮挡的边界建立一个分层（hierarchical ）的分割，并且使用不同的种子点和参数来切割图产生区域。产生的 使用大量的线索和鼓励多样性的角度排序。

• Rigor (GC) [28]：是 CPMC 的一个改进，使用多个图切割和快速的边缘检测子来加快计算速度。

• Geodesic (EC) [22]：首先使用 [36] 对图片过分割。分类器用来为一个测地距离变换标定种子点。每个距离转换的水平集（Level sets）定义了（figure/ground）的分割。

• MCG (EC) [23]：基于 [36]， 提出一个快速的用于计算多尺度（multi-scale）层次分割进程。使用边缘强度来合并区域，生成的目标假设（object hypotheses ）使用类似于尺度，位置，形状和边缘强度的线索来排序。首先用现成方法快速得到一个层次分割的结果，然后利用边缘信息进行聚合。

Window scoring proposal methods

不同于前者需要通过聚合小块来生成候选框，这里的方法是先生成候选框，然后直接打分排序来过滤掉低分的候选框。介绍两种比较出名的方法，与 grouping approaches 比，这些方法值返回边界框（bounding boxes），因此速度更快。但是，除非它们的窗口采样密度很高，否则这些方法位置精度很低。

• Objectness [12], [24]：最为最早和最广泛的一种 proposal 方法。它通过选择一副图片中的显著性位置作为 proposal，接着通过颜色，边缘，位置，尺寸，和 superpixel straddling 等多个线索对这些 proposal 打分。

• Rahtu [25]：以 一个包含采样区域（单个，两个和三个超像素）和 多个随机采样的框的大的 proposal 池作为开始。采用类似于 Objectness 的打分策略，但是有些提高 （[40]添加了额外的 low-level features 和 强调了恰当调优的非最大抑制（properly tuned nonmaximum suppression）的重要性）。

• Bing† [18]：通过边缘训练一个简单的线性分类器，并且以一个滑动窗口的方式运行。使用充足的近似，获得一个非常快的类未知的检测子 （CUP中每帧 1ms）。CrackingBing [41]表明一个有很小影响和类似性能的分类器可以通过不用查看图片的方式来获得 （分类性能不是来自于学习而是几何学）。在CPU上能够达到ms级别。但是被文献[40]攻击说分类性能不是来自于学习而是几何学。

• EdgeBoxes† EC [20] ：基于目标边界估计（通过 structured decision forests [36], [42]获得）形成一个粗糙的滑动窗口模式作为开始，使用一个后续的 refinement 步骤来提高位置精度。不学习参数。作者提出通过调节滑动窗口模式的密度和和非最大抑制的阈值来调优方法用于不同的重叠阈值。跟selective
search一样是一个不需要学习的方法，结合滑窗，通过计算窗口内边缘个数进行打分，最后排序。

• Feng [43] ：通过搜索显著性图片内容来找到 proposal ，提出了一种新的显著性度量，包括一个潜在的目标能被图片的剩余部分组成。它采用滑动窗口模式，并通过显著性线索对每个位置打分。

• Zhang [44] ：提出在简单的梯度特征上训练一个级联的排序 SVMs。第一阶段对不同的尺度和长宽比（aspect ratio）训练不同的分类器；第二阶段对所有获得的proposals 排序。所有的 SVMs 使用结构性的输出，对含有更多目标重叠的窗口打分更高。因为级联在同样的类别上训练和测试，因此不太清楚它的泛化能力。

• RandomizedSeeds [45] ：使用多个随机的 SEED 超像素映射图 对每个候选窗口打分。打分策略类似于 Objectness 的 superpixel straddling （没有额外添加的信息）。作者展示使用多个超像素映射（superpixel maps ）可以明显地提高召回率。

Aliternate proposal methods
ShapeSharing [47] ：是一个无参的数据驱动的方法，通过匹配边转换目标形状从范例（exemplars）到测试图片。生成的区域使用图切割合并和提纯。 Multibox [9], [48] ：目前笔者所知唯一基于CNN提取proposal的方法，通过CNN回归N个候选框的位置并进行打分，目前在ImageNet的dectection
track上应该是第一的，训练一个神经网络来直接回归一定数量的 proposals （不需要在图片上滑动网络）。每个 proposals 都有它自己的位置误差 。该方法在 ImageNet 表现出最好的结果。
Baseline proposal methods

这里用了Uniform，Gaussian，Sliding Window和Superpixels作为baseline，不是重点就不展开说了。

各种OP方法对于复现的鲁棒性的讨论

这里作者提出这样的假设：一个好的OP方法应该具有比较好的复现能力，也就是相似的图片中检索出来的object应该是具有一致性的。验证的方法是对PASCAL的图片做了各种扰动（如Figure 2），然后看是否还能检测出来相同的object的recall是多少，根据IoU的严格与否能够得到一条曲线，最后计算曲线下面积得到repeatability。
这里图表很多具体请看原论文，这里直接上作者的结论，Bing和Edgeboxes在repeatability上表现最好。

各种OP方法的recall

这里提出了好的OP方法应该有着较高的recall，不然就要漏掉检测的物体了。这里讨论了三种衡量recall的方式：
Recall versus IoU threshold: 固定proposal数量，根据不同的IoU标准来计算recallRecall versus number of proposal windows: 跟1互补，这里先固定IoU，根据不同的proposal数目来计算recallAverage recall(AR): 作者提出的，这里只是根据不同的proposal数目，计算IoU在0.5到1之间Recall。
数据集方面，作者在PASCAL VOC07和ImagNet Detection dataset上面做了测试。
这里又有不少图，这里只贴一张AP的，其他请参考原论文咯。
还是直接上结论
MCG， EdgeBox，SelectiveSearch, Rigor和Geodesic在不同proposal数目下表现都不错如果只限制小于1000的proposal，MCG,endres和CPMC效果最好如果一开始没有较好地定位好候选框的位置，随着IoU标准严格，recall会下降比较快的包括了Bing, Rahtu, Objectness和Edgeboxes。其中Bing下降尤为明显。在AR这个标准下，MCG表现稳定；Endres和Edgeboxes在较少proposal时候表现比较好，当允许有较多的proposal时候，Rigor和SelectiveSearch的表现会比其他要好。PASCAL和ImageNet上，各个OP方法都是比较相似的，这说明了这些OP方法的泛化性能都不错。
各种OP方法在实际做detection任务时候的效果

这里作者在OP之后接上了两种在detection上很出名的detector来进行测试，一个是文献[54]的LM-LLDA（一个DPM变种），另外一个自然是R-CNN了，值得注意的是，这两个detector的作者都是rbg。。。真大神也。。。
这里用了各种OP方法提取了1k个proposal，之后作比较。
也是直接给作者结论：
如果OP方法定位越准确，那么对分类器帮助会越大，因为定位越准确，分类器返回的分数会越高：在LM-LLDA和R-CNN下，使得mAP最高的前5个OP方法都是MCG,SeletiveSearch,EdgeBoxes,Rigor和Geodesic。
分数一览如下图。通过分析，作者发现AR和mAP有着很强的相关性：作者用AR作为指导去tuning EdgeBoxes的参数，然后取得了更好的mAP（提高1.7个点）
全文的总结和讨论

总结：
对于repeatability这个标准，目前的OP方法效果都一般。可能通过对噪声和扰动更加鲁棒的特征能够提高OP方法的repeatablilty。但是repeatability低不代表最后mAP就低，比如SelectiveSearch，所以最后还是看要应用场景。如果OP方法定位越准确，那么对分类器帮助会越大。所以对于OP方法来说，IoU为0.5的recall不是一个好的标准。高recall但是定位不准确，会伤害到最后的mAPMCG,SeletiveSearch,EdgeBoxes,Rigor和Geodesic是目前表现最好的5个方法，其中速度以EdgeBoxes和Geodesic为优。目前的OP方法在VOC07和ImageNet的表现都差不多，说明它们都有着不错的泛化性能。
讨论：

结果是： 1. Scale：影响较大，bing 更好点 2. JPEG artefacts：bing 最好 3. Rotation :都差不多 4. illumination：趋势相似，bing更好 5. Blur：相似 6. salt and pepper noise：该因素影响比较大 结论：Bing在repeatability中表现最好，EdgeBoxes也不错。

参考博客：
http://blog.csdn.net/shanglianlm/article/details/46786303


展开全文
• What makes a good model of natural images
• no.16 W3D4 Hello, everybody. Try to imagine this situation, what if I bring a well-made robot ... What if this robot can dance, sing and even cook a delicious meal for you? Would every one of you ...
no.16
W3D4
Hello, everybody. Try to imagine this situation, what if I bring a well-made robot here? What if this robot can dance, sing and even cook a delicious meal for you? Would every one of you show greatest interest in this fantastic big toy?  … Yes, I believe you do. All of you like high-tech machines. Most of you ask your parents to buy you a laptop after college entrance exam, some of you may possess an iphone and use it to blog and surf the Internet frequently during your break between classes. To boys, model plane is also a common dream when you are little. But with which on earth could engineers make so many magnificent machines and even let them move?

Today, I will tell you how they make it, with showing you the three main factors for engineers to design a product. Our emphasis will be laid on mechanical structure, electric and simulation in the following.

First of all, let’s talk about mechanical structure. Structure is to machine, as the body to a human being. As the first but foremost step of manufacturing, designing suitable structure helps to support the weight of a machine, strengthen their quality and make a product good-looking. What’s more, mechanical structures make possible the specific movement of a product. Gear, for example, does a lot help for the transmission in a typical machine.  Up to now, there have been thousands of millions of different but ingenious mechanical structures invented by those talents in this field. It is those epoch-making mechanical structures that push human civilization into a glorious big time.

When it comes to electric, it’s like blood, which connects all the organs in a human body. Why would I say that? Let me tell you the reason. We know motor drives the machine to move. But what drives motor to move? The answer is electricity. Electricity flows throughout the machine and makes every “organ” alive. What a wonderful thing! Generally speaking, electric in machine designing includes circuit diagram, electronic components and electronic chip. Engineers draw circuit diagram, write programs into chips, and after that they put many components altogether to ensure the final situation follow their original designing.

After we understand the importance of structure and electric, let’s talk about simulation. Simulation is a certain particular process to imitate a model in virtual world. In the modern time, creating inventions in the computer and testing it after that definitely shortens the circle we manufacture a certain machine. Besides that, analysis to the virtual products is also available these days. Obviously, it is simpler to test a model in computer rather than in the real world. It will save a lot of money, right?  All in all, what an essential and economic way simulation is and without it, we nearly can’t do anything effective today to create a product high in quality.

Now, you can see that structure, electric and simulation show their own essential role in the manufacturing. Works by engineer are not a secret job as we expected before.  It’s just a process of applying one technique into making a product. That’s it! Wish all of you could enjoy it.
Thank you.

Jack
转载于:https://www.cnblogs.com/tianya-lemon/p/5294029.html
展开全文
• ## Whatmakes a good leader

千次阅读 2011-07-13 09:02:15
What makes a good leader --- from employees' standpoint of view. 1. Control anger. Anger cause more problem than they solve. 2. Do not sho
What makes a good leader --- from employees' standpoint of view.  1. Control anger. Anger cause more problem than they solve.  2. Do not show off your power  3. Know every solder's strength and use them  4. Unit your team  5. Make your team envy you  6. Able to let go  7. Take the responsibility
展开全文
• What makes a popular mobile game? 1. Background The mobile games industry is developing so fast, with companies spending vast amounts of money on the development and marketing of these games to an eq....
• <div><p>I am currently working with the TS3ClientTests file and wondering what OnErrorEvent outputs to me, it does not show that the client is flooding</p><p>该提问来源于开源项目：Splamy/TS3...
•  Do not waste time to put in what was left out of a person.  Try to draw out what was left in a person.  That is hard enough.  This is the foundation for their success. Revolutionary Insignts  ...
• Introduction We all want to be good programmers. To become a good programmer you have to know what makes a good programmer. The following is a list of things that in my experience make a good program
• 奇人奇文，图像是否可以被记忆？又是什么造成了图像的记忆呢？作者的探讨非常值得思考
• 目标识别What Makes Paris Look like Paris源码.
• <p>To that end, I was trying to find what makes a plug-in <em>awesome</em>? Are there any requirements for plugins added to the list that I could learn from and use to improve my work, or is it purely...
• <div><p>Lively <a href="https://18f.slack.com/archives/wg-code-reviews/p1445377295000008">discussion</a> in chat; let's open 'er up.</p><p>该提问来源于开源项目：18F/development-guide</p>...
• What Makes Node.js Faster Than Java?
• Being volatile in price is the major characteristic of the crytocurrency, but what makes the market unstable? This article is going to briefly investigate factors that may arouse price move in market....
• 转自：http://joshsymonds.com/blog/2013/11/03/what-makes-a-good-programmer-good/ I’ve worked with a lot of programmers over the years – some of them super amazing, and some distinctly lackluster...
• There’s a lot of talk these days about what makes a good “culture”, whether you’re an engineer, a software developer, or a chef. It’s all about finding a work environment that not only is ...
• 下面是我给公司期刊写点一篇文章，...[b]What Makes Us Unique and Different[/b] We know Microsoft by Windows and Office, we know Google by its search engine, and we know Apple by Mac and iPod. I alway...
• If people were honest about what makes great teams in an organization, they'd say part of it is the emotional connections that allow both openness and caring."   Diverse range of talents. ...
•  What makes you different,makes you beautiful  是什么让你与众不同，让你如此美丽  What's there inside you shines through to me  你的内心深处 一直照射到我  In your eyes I see all the love I'll ...
• d like to write on this topic, I intend to cover all the important aspects that makes HTML5 the future of websphere. Please let me know your valued feedback, so I may start writing. <p>Thanks</p><p>...
• OLAP vs OLTP: what makes the difference OLPT and OLAP are complementingtechnologies. You can't live without OLTP: it runs your business dayby day. So, using getting strategic information from OLTP is
• TITLE: What makes ImageNet good for transfer learning?

...