Persuade or persevere? Computer Vision Industry Overview
1 From hot to controversial
Computer Vision (CV for short) is a discipline that studies how to enable computers to obtain information from images or image sequences and
understand their information. Its main purpose is to extract descriptions of the world from images or image sequences. From an engineering point of view, it
studies how to use algorithms to simulate the human visual system, so as to complete a series of tasks that humans can complete through vision. The most widely known
application is "face recognition".
As one of the fastest-growing fields related to deep learning in the past two decades, computer vision has made countless technical people dream and fly
away. There are two main reasons why people are chasing computer vision: First, this field is very interesting and has the potential to change the world.
80% of the information obtained by the human brain depends on the formation of vision, and the economic and social benefits brought by the successful simulation of the human visual system are immeasurable.
From a scientific point of view, computer vision is worth a lifetime of research for researchers. Entering this field, you will have the opportunity to have your own
career (career), not just a job (job). Second, at the beginning of the 21st century, the field of computer vision has achieved rapid development and
has a huge number of practical landing scenarios, which means that this field has industrial demand, academic potential, and most importantly, a high
economic value. These factors make computer vision a favorite in the capital market, and the price of talents in the industry has also risen. Computer vision has become the
career development direction that many people dream of.
But soon, people are staying away from computer vision. Not to mention that deep learning itself is a subject with a certain threshold. In
the fall of 2019, the recruitment of algorithm posts also saw a scene of "Twilight of the Gods". CV positions
have almost formed a complete seller's market. Since then, people believe that the field of computer vision has been seriously involved and academic research has stagnated. Although there are many industrial
landing scenarios, the cost is huge. It seems bright and bright, but the cost performance is not high. All of a sudden,
a scene of "great escape" from computer vision was formed. Many graduates with AI dreams turned to development positions, gave up computer vision, and even completely gave up the road of AI.
From hot to full of controversy, computer vision has only gone through a few short years, which is related to the rise and fall of the global AI trend and people's heightened
Being optimistic about the field of computer vision has a lot to do with it, but it is more caused by the academic characteristics of low threshold, high upper limit, and steep learning curve in the field of vision
.
2 Low threshold, high upper limit, steep learning curve
Deep learning is a technology with a threshold. Anyone who has a little understanding of the concepts of "algorithms" and "artificial intelligence" dare not easily say that deep learning, especially
computer vision, has a low threshold. But in the vision industry, compared to reaching the level of "familiarity" or "proficiency", getting started with vision is indeed
too easy-first of all, everyone knows that the core of computer vision is the convolutional neural network CNN, so learn CNN first That's right, usually an
excellent teacher only needs 30 minutes to give people a general impression of the workflow of the convolutional neural network. The rest is to
find a piece of code on github or even CSDN, Baidu, After a few days of tinkering and running the code, it is considered to have completed the first
"vision system neural network" in life. For more difficult models, the results can also be easily obtained by using the package adjustment method. A few people will find a few
examples of image recognition to complete their own learning, but most people only stop at running through (other people's) code, and then move on to the next field, which is not
difficult.
But learning in this way is still a long way from becoming a "computer vision talent". Many people are dumbfounded after completing the study of classic models, and
will find that they "have no idea where to go further", because there are only a few classic models in this field, and it seems that there is no more
content to learn ( Learning algorithms, besides learning models, what else can you learn?). But looking at the naked "top paper"
requirement in the recruitment conditions, I can't imagine where the paper should come from. I can only start from the perspective of "which field is good for paper". At this point, you have
reached the bottleneck of computer vision advancement. The master is in the atmosphere, and you are on the 18th floor underground, but you can't see where the difference between you comes from
. Even if you have worked hard to find the difference between you, you will find that you can't directly step from the simple CNN architecture into the
realm of God that "reading documents, writing documents, and looking at formulas can reproduce them". More than 90% of learners are unable to cross this steep learning curve and continue to move forward. It
is not surprising that computer vision is too demanding and has serious introversion.
Why is there such a situation in the visual field? There are two root causes: