Due to inconvenient to put links, a better reading experience, please see: How Automation turn SLAM or machine learning post?
This article from the issue of the same name known on the peace, the original link:
https://www.zhihu.com/question/266685012/answer/336327001
title Lord is a Beijing Automation 985 universities of undergraduate and postgraduate (Master control engineering), just graduated about half a year, the first job was in a small entrepreneurial company to do the robot control engineer. Work content is very complex, mainly the work of some stm32 development and signal processing and the like, chaotic management of the company, and the company's lack of experience in the old rich technical staff, just graduated as a newcomer, I feel completely learn anything, lack of growth, and the distribution of work is not their own interest, and I worked for less than six months to resign, and intend to take advantage of the New Year ready to look for a reliable point of work. Mainly want to slam the machine direction and the direction of learning.
My own situation itself is this:
Project experience:
no internship experience, basically their own projects in the laboratory and machine learning and slam it does not matter, the main tune PID what is pure engineering tasks. Undergraduate complete set is made by the signal recognition SVM, but undergraduate complete set is very simple.
Personal knowledge base:
1, machine learning this I have been in school from undergraduate, Andrew Ng's class read the front half, "statistical learning methods" all read it again, which the algorithm is also largely pushed through again. Zhou Zhihua watermelon also read the book. But there is no actual project experience (do not know the Titanic survivors forecast and TensorFlow Digital Recognition count).
2, slam this one I am self-taught "visual fourteen speak slam" a book, principles are read, about 70% of the code in the book have their own hand struck again, orb-slam2 papers read the code carefully read.
Please workplace old drivers slam and machine learning industry's answer to my question:
the case I do not have internship experience and project support, turn to slam or machine learning jobs, whether it can be accepted by the company? Or that I want to achieve what extent, in order to be accepted related companies? The company will be willing to accept the new culture from zero it? (I count from 0.5 bar)
Workplace Can you give some of the old drivers to switch actionable advice?
In this paper, several excellent answers were consolidated, interpretation A master of all. If infringement, please contact delete
the following views do not represent the life of computer vision point of view, only as a reference.
Half lay idle:
Automation is the next graduate, it does not seem there is a switch that controls automation to engage in, engage in machine learning or slam normal.
Machine Learning wide range of applications, low threshold, currently learning materials is very rich, it is not hard to learn. In contrast lot less slam materials, books also are voluminous book, learn it was quite time-consuming. Behind slam involves mathematical knowledge more widely, we can find out exactly what it is not easy (deep learning in addition there are other mathematical gradient descent do).
slam post currently seems unlikely recruit to the human stage. Although some people think that Slam has been resolved in theory, but the reality is:
most people slam write a program from scratch, most write less than the level of current open source program;
the case of open-source program does not meet the demand, most do not change (not in-depth understanding of the program);
SLAM applications are mainly in the automatic pilot, robot, AR / VR. Domestic autopilot very active in recent years, BAT and each has a depot layout, a little-known start-up companies can count more than 30. Robot, then the logistics needs of the park is also more obvious, Jingdong, rookie, and many start-up companies are doing, ARVR not to say, content and ecological Can it still a problem. In case you have a master's degree, it should not be difficult to find the corresponding positions. If you can write a complete set of slam, or have a very deep understanding (see to understand, to change the dynamic) to open a program, most companies should be you (as far as I understand many of the company's staff to slam Learn also the depth). If you have a solid theoretical foundation and then some of the better.
As a suggestion, you can try to write a simple slam project, on kitti, euroc and other data sets and open source program after comparison test, the results put your resume, and you will find the right job increases the probability of a lot.
Pickles Husky：
Doctoral SLAM direction, more recently, to find a job, simply say a few words.
First clarify that, SLAM and machine learning is absolutely not an order of magnitude of the field. Machine learning for many companies and businesses are very basic tool, and this trend is increasingly evident, after a lot of jobs will need to master the technology of machine learning. Machine learning with many fields cross, such as SLAM place now in the application of machine learning more and more. Bluntly said, to grasp the importance of machine learning and data processing skills, equivalent to master important in the field of programming language, such as C ++.
Narrow multi-application surface of SLAM. SLAM is now the main demand of the robot, autopilot and ARVR field, if SLAM is your focus, you are limited employment opportunities in these sectors.
And to say it bluntly that, even in these areas, SLAM nor drive the core technology for these applications, SLAM only as a provider of localization and mapping of sub-modules only. Currently, many applications SLAM problem has not been completely resolved, we spend a lot of energy to solve the problem on SLAM; and once SLAM problem a better solution, SLAM will only provide positioning system and map building tool only, everyone's attention will more on the more important issues, such as autopilot in perception, semantic information and behavior prediction problem, manipulation problems in the field of robotics, and so on. Although I do SLAM, but SLAM irresponsible to predict the field of heat is difficult to continue for a long time, at least the next ten years in order to view the order of the decline is inevitable. The machine learning and data science the opportunity now and future potential are difficult to limit.
SLAM complete discussion of the industry, and then look at the background of the main problems. Features on the SLAM and machine learning to master the skills a bit like, is dominant in this area is relatively low threshold, but invisible threshold is relatively high. opensource down some of the code we can run down a better demo, like machine learning which we follow the tutorial to run mnist very simple example of the same. But things did not opensource application background, a certain proportion in a real scenario, even a large area of work is not very common. When you really get to business which, companies need you to do SLAM module can operate stably in the application scenario, this time either on your own from zero to write, or change opensource, you need to have a good understanding of the SLAM principles and systems, to adjust the system to deal with cases of a number of fail. Enterprises can deal with this problem, a couple of years of actual project experience coupled with a considerable amount of reading (MVG and classic tomes such as paper, at least in the traditional paper) is a must.
I am very optimistic about the main issues on their own practice at home. SLAM project can do at home are generally not the industry background (this is even worse condition than the machine learning, machine learning, so at least there is kaggle "close to" the industry can do a project, in addition to kitti SLAM and other industry completely detached), the industry is very practical value related to the project background. And willingness to train a new business from scratch is very low, you're the first company to recruit hope you have actual output, rather than spend money to train you; Secondly, as said before training SLAM direction or partial long time.
If the main problem bent on going to line SLAM skills as the core, there are two possible paths to consider the main issues: 1. Dr. Master who read the relevant direction, after all, the probability of zero to recruit graduate students based on much higher than enterprise. 2. take the development of the line, you need to question the master has a strong C ++ background (SLAM developed in C ++ is now absolutely mainstream), and software development of identity into related companies SLAM team conducted SLAM software in experienced people to guide development, which is the only channel I can think of to learn SLAM in the enterprise inside it. . .
CLEMENT HUANG：
Our company was looking for an automated background, willing to learn SLAM, strong programming capabilities engineer. Coordinate Shanghai. So see this issue, I could not help to answer a few.
First introduced himself, automation professional undergraduate and postgraduate schools are 985, he spent three years in state-owned enterprises, foreign invested enterprises for 12 years, Venture 2 and a half years. So for the control majors confusion should be very personal experience of. Control of a professional thing students learn a lot, but to always have a job in a powerful does not come out feeling, the key is the application of automatic control is too broad, motion control, drag, process control, computer theory, DSP or be Soc (System on Chip), power electronics, a variety of programming languages, software engineering, but also to the graduate school of digital signal processing, control various optimization algorithms, and even co-ordinate learning, including robotics recent fire up, SLAM, artificial intelligence, game theory . You said a school student in school or just graduated three years to get new people to go touch the door, the industry is just a small obstacle can stop living a confident young man, let alone come up with a single if weapons, but we are also doing other professional. Such as electric power electronic circuits, certainly dry, but to learn the power system, software dry, but computer algorithm did not use strong math, engage in communications and telecommunications projects there are gaps. Even motor control, to know that colleges and universities also have special electrical engineering. Control students may learn a year's electric drive motor have not seen a real, but can not tell DC, AC, DC brushless, permanent magnet, AC synchronous. However, I have seen countless automated Daniel, later in the direction selected (non-automated) have become the most outstanding dry, it was even three years throughout his career began to spend more time, effort, countless injuries brain cells only achievements.
Back to the topic, I can give advice is to find a platform for their own direction and maintain strong desire to learn, three-year rolling down the dry, short-term vocational window period is acceptable, but more than six months or more, the basic on difficult to get an interview, the company is evaluating a candidate for such a risk is too high. And you learn and learn on the job at home is completely different, read many books does not mean you have into the line.
Do not give yourself too narrow positioning algorithm of real people in a company are extremely rare, the algorithm needs to use its own code to achieve it. To test it will be able to understand the underlying interfaces to your right. But also to understand some of the industry background applications, and I can only say that learning is just beginning.
Liu Chang:
A small first know almost transparent, I would like to take your treasure wrote to enter new areas or SLAM SLAM hesitant to enter the field of graduate student (? Note, graduate students, graduate students do SLAM Why because in fact the time was quite tight) some advice:
1 . can not touch the robot does not touch physical entities robot
2. find a good part started
because SLAM too broad, deep content is also involved in the school is a place where innovation requirements, so read fourteen speaking (it has been ring after the inside called SLAM Bible, thanks to high lighthouse), ran most of the code book, do not say, ah ~ ~ so excited I have all mastered SLAM friends - but calm down and think about it, I would like to study which part is based on the front-end multi-view geometry? Still more interested in optimization problems, we want to optimize the probability of two books from the robot and convex start? Or is what I want to do some machine learning loop? Or that I want to abandon the existing set of programs directly to a deep learning of end-VO (this has been a relatively well-known universities do, but want to do in-depth studies should also be right), with do research problem, the problem of abstraction. If you just want to optimize positioning accuracy down, really do not have to find a physical robot, learn to use the data set to test their own programs, will focus more on methods and optimized C ++ (especially for the main answer this could have been mechanical, programming is C ++ primer chapter level ~). today to answer them right, too late, after a time and then answer.
Slamer:
I also happen to automation, a robot who just graduated into the company, to do work related to positioning mapping, machine learning little understanding, mainly to talk about slam. Against this background
slam direction is the combination of engineering and scientific disciplines, under the premise of good mathematical foundation also needs a strong engineering (especially c ++) capability. If you want to engage slam title the main direction, first of all you can first understand the more popular slam solutions, read the relevant open source projects, such as laser slam in gmapping, cartographer, open source projects not only allows you to have a macro understanding of the slam, but also be able to learn good coding style and code techniques. Secondly, learning the theory of knowledge, like the nonlinear least-squares Kalman filter, particle filter, numerical optimization, automatic derivation, etc. With this foundation, slam entry has only just begun. If you want advanced, try to combine theoretical knowledge he learned to write a slam, this time for these theories have a deeper understanding of written works at the same time, you may find some things you can not book high school, such as parameter adjustment, control parameter adjustment skills, good parameters like cosmetic surgery, the robot can make a very good work, keep up the pace tight behind the times, read more slam-related paper. Finally, do not be impatient slam, slam there will be a lot of small problems in engineering, maintain a good attitude, I wish you good luck.
Welcome to public concern number: computer vision, life, explore a new world with computer vision - has long been an advantage, learning should not go it alone, there are tutorial materials, work practice, answering questions such as, high-quality learning circle to help you avoid detours, Quick Start !