秉持尊重学生和家长隐私的原则,以下所展示的图片,已经去掉隐私及涉及知识产权的信息。
还有N多学术素材,由于知识产权或家长原因,不便公开,敬请理解。
本文作者GZ同学,品学兼优,目前是985大学二年级学生。
GZ同学有更高的梦想,希望本科毕业后能够到美国最好的大学读博士。为此,参与了美国名校科研,增加学术背景,开拓视野,获得真知。
大一、大二的学生参与科研,会有更充裕的时间与大学老师进行交流,学生表现好,就会成为科研组和实验室的一员,下一个假期,学生依然可以继续完成科研任务,完成更有难度的学术挑战。
本文是学生在美国大学科研学习结束后所写的感受,供学生及家长参考。
注:中文译文是编辑进行的整理,英语能力较好的学生和家长建议直接阅读英文。
After a month long research experience at American university, I felt my horizon is broadened and I did experienced an unprecedented academic experience in my life. I hope my experience could be a powerful boost for the future research in the area of artificial intelligence & computer vision. What I learned from American university is not only the process, methods of algorithm improving, but also the spiritual belief of university professors.
在美国的大学学习一个月后,我感觉我的视野开阔了,我的确经历了一生中前所未有的学术经历。我希望我的经验可以为未来的人工智能和计算机视觉领域的研究提供有力的推动。我从美国大学学到的不仅是改进算法的过程、方法,而且是大学教授的精神信仰。
My research experience started on July 16th. The first thing that surprised me was the self-introduction part in the first meeting. We exchanged views and suggestions on each others’ self-introduction. Although this sharing part was done at the teacher's request, but I must say that this is the fundamental requirement of the adaptation to the month long group research life. During this month, communication is the top priority for any help and suggestions are provided from the group discussion.If you felt shy and did not talk with each other, then you would get nothing in this month.
我的研究经验从7月16日开始。让我吃惊的第一件事是第一次会议上的自我介绍部分。我们就彼此的自我介绍交换了意见和建议。虽然这部分分享是应老师的要求而做的,但我必须说,这是适应长达一个月的团队研究生活的基本要求。在这个月里,交流是所有帮助的首要任务,小组讨论会提供建议。如果你感到害羞,没有互相交谈,那么这个月你将一无所获。
At the beginning, I hope to do a bit of real tough research. So I chose the join the algorithm group which is basically focus on the activity & image understanding. My topic is the tiny falling object detection & similarity comparison of intelligent monitoring (the previous one is the subtopic of the latter). I took a few days finished the basic requirement from professor N --- to track the bigger object in the video and fitting the falling trajectory. From then on, I was able to conduct a real research on the tiny object (pixel level in the image) detection, which is a area still has no satisfactory solution.
开始时,我希望做一些真正的艰苦研究。所以我选择了加入算法组,基本上集中在活动和图像理解上。本课题是智能监控中的微小落体检测与相似性比较(前者是后者的子课题)。我花了几天时间完成了N教授的基本要求——跟踪视频中较大的物体并拟合下降轨迹。从那时起,我就能够对微小物体(图像中的像素级)的检测进行实际的研究,而这个领域仍然没有令人满意的解决方案。
So, after the basic part, I spent my first week on completing a decent thesis proposal. First, I read a few related papers (about Fast-R-CNN, optical flow, SPP-Net and etc.). At the same time, I have several discussion with professor N according to the information I got from paper. Professor N's answer made me convinced that the topic I chose was still a valuable and had no satisfactory solution. I think my research purpose can be set as: Accurately locate falling objects based on the influence of cell monitoring; Predict the whereabouts, original point and time based on the falling trajectory;Branch direction: Detect if the throwing action throws an object. I finished two basic program (optical flow & inner-frame difference method) in order to test the algorithm’s effect on the video I got.
因此,在基本部分之后,我花了第一周完成了一份体面的论文提案。首先,我读了一些相关论文(关于快速RCNN,光流,SPP网等)。同时,根据我从论文中得到的信息,我与N教授进行了几次讨论。N教授的回答让我确信,我选择的话题仍然是有价值的,没有令人满意的解决方案。我认为我的研究目的可以是:根据细胞监测的影响精确定位落体;根据落下轨迹预测下落、原始点和时间;分支方向:检测投掷动作是否抛出物体。完成了两个基本程序(光流和帧内差分法),测试算法对视频的影响。
Since first week's program was only able to initially detect the bigger falling objects.I spent three days modifying my program in order to detect a clearer range of the falling objects. I added some specific noise reduction processing to get a better result. At the same time, I determined the purpose of my research in the study of tiny falling objects such as empty bottles, empty can, even cigarette ends. From the result I got we could find that some bigger object could be detected clearly. We could see the edge in the picture without background. But some smaller object like cigarette ends could not be detected. At the same time a new problem has emerged.Because I have enhanced the accuracy of the program. The current program can detect the tiny jitter of the camera, which means the current program introduces a fatal error. All the pattern noises could easily covered the trace of the tiny object. I am still thinking about how to solve this problem. Professor N suggested me to learn the algorithm called Eulerian video magnification. It’s a method that could amplify the subtle motion like heart beat in the video. Maybe I can apply this algorithm slightly improved to the magnification of the falling object’s movement.
因为第一周的程序只能初步探测到较大的坠落物体。我花了三天时间修改程序,以便探测到更清晰的坠落物体范围。我添加了一些特定的降噪处理,以获得更好的结果。同时,我决定了我的研究的目的,在研究微小的下落物体,如空瓶子,空罐子,甚至香烟头。从得到的结果我们可以发现一些较大的物体可以被清楚地检测到。我们可以看到图片中没有背景的边缘。但一些较小的物体,如香烟端,不能被检测到。同时,出现了一个新的问题,因为我已经提高了程序的准确性。当前程序可以检测相机的微小抖动,这意味着当前程序引入了致命错误。所有的图案噪声都能很容易地覆盖微小物体的痕迹。我仍在思考如何解决这个问题。N教授建议我学习欧拉视频放大算法。这是一种可以放大像视频中心跳的微妙运动的方法。也许我可以把这个算法应用到落体运动放大率上。
In the third week’s experiment, I have summarized several problems that I have met during the tiny object detection research. During this week, I mainly focused on one of the problem that the less feather point could be detected and they are easily influence by the patter noises. In the area of video magnification, there are two main algorithm “ lagrangian video magnification” & “eulerian video magnification” I choose the second algorithm because there are very few feature points available for extraction in my data. I used laplacian pyramid to enhance the amplification. After finishing the program, I checked it on the data set given by the paper ” Eulerian video magnification for revealing subtle changes in the world”
在第三周的实验中,我总结了我在微小物体检测研究中遇到的几个问题。在本周中,我主要关注一个问题,即羽毛点较少,并且它们很容易受到模式噪声的影响。在视频放大方面,有两种主要的算法“拉格朗日视频放大”和“欧拉视频放大”,我选择第二种算法,因为在我的数据中只有很少的特征点可用于提取。我使用拉普拉斯金字塔来增强放大。完成这个程序后,我在论文《揭示世界微妙变化的欧拉视频放大率》给出的数据集上检查了它。
Here are the amplified video got from my program. The baby’s breath and his subtle check movement is amplified so we could see it easily. But the problem is that the the color of the background showed severe distortion in the first two seconds. I am still trying to solve this problem.Another problem is that the program I wrote takes up too much system memory. I still need to improve the performance of the program.
这是从我的节目中得到的放大视频。婴儿的呼吸和细微的检查动作被放大,所以我们可以很容易地看到它。但问题是背景的颜色在前两秒显示出严重失真。我还在努力解决这个问题。另外一个问题是我写的程序占用了太多的系统内存。我仍然需要提高程序的性能。
I finished my amplifying program in the last week. The result I got from the test video (baby video) is encouraging. I tried the program on the falling object data set. This time, the result is not surprisingly not that satisfactory. Although I could amplify the bigger falling object in the video, I still could not find any trace of the tiny object’s trajectory in the cigarette video. Before leaving, I collected everything I found into a report to communicate with professor N. He gave me a suggestion that I could check the method in his paper to create a physical model to solve this topic. He also affirmed my idea:
use upsample instead of downsample to provide more features to the algorithm
try the frequency image to find the connection between tiny object and the frequency( use FFT in three-dimensional area).
上星期我完成了我的放大计划。我从测试视频(婴儿视频)得到的结果是令人鼓舞的。我尝试了关于坠落对象数据集的程序。这一次,结果并不出人意料,不令人满意。虽然我可以放大视频中较大的坠落物体,但我仍然无法在香烟视频中找到任何微小物体的轨迹。在离开之前,我收集了我所发现的一切,与N教授交流。他给了我一个建议,我可以检查他的论文中的方法来创建一个物理模型来解决这个话题。他也肯定了我的想法:
使用UpStor代替下采样来为算法提供更多的特征
尝试频率图像找出微小物体与频率之间的联系(在三维区域中使用FFT)。
In addition to this valuable research experience, the conversation with professor N also help me a lot. He told me that all the American professor had the perseverance to explore new areas but not just simply combined algorithms. It’s because some people are willing to go to the hard bones that new algorithms could be discovered. In fact, Chinese and the Americans now have little difference in their lives and even most of their ideas. But this kind of spirit of exploration is what I think is the most lacking spirit in the country.
除了这个宝贵的研究经验,与N教授的谈话也帮助了我很多。她告诉我,所有的美国的大学的教授都坚持不懈地探索新的领域,而不仅仅是简单的组合算法。这是因为一些人愿意去探索新的算法。事实上,中国人和美国人在他们的生活,甚至他们的大多数想法上几乎没有差别。但这种探索精神是我国最缺乏的精神。
北京博师屯儿教育科技有限公司(博士屯教育),注册在北京,团队管理者平均行业经验8年。专注美国学界合作,目前合作大学遍布全美,深度合作区域有波士顿区域,纽约区域,旧金山区域等。
博士屯教育,专注美国名校合作,为学生定制有利于成长的一站式背景提升方案,助力本硕博名校录取。
博士屯,就是做背景提升的,目标就是提高本、硕、博录取质量。
As the most energetic provider of research educational services in China and North America, Ph.D village offers education for helping students to get into better college and helping improve their quality of life. Our wide range of educational programs, services and products includes research program, elite program, visiting scholar program, as well as paper online education.
Ph.D village has become one of the most recognized brand in research education area.