Shulou(Shulou.com)06/02 Report--

With the rise of autopilot "tuyere" in the past few years, more and more capital and R & D have been invested in the field of autopilot.

Relevant agencies predict that the market potential of semi-self-driving and fully self-driving cars is huge in the coming decades. By 2035, there will be about 8.6 million self-driving cars in China alone, of which about 3.4 million will be fully autonomous and 5.2 million semi-autonomous.

However, autopilot is a very complex engineering system, which requires the integration of many technologies and precision cooperation, and it is impossible to rely on the power of capital to break out rapidly in a short time, and the commercialization of autopilot is also faced with challenges in all aspects. So for a long time, autopilot gives the impression that it is "very hot but far away".

There are many reasons why it is difficult to land on autopilot, one of the core factors is the imperfection of autopilot perception system.

In self-driving technology, perception is the most basic part. Without quantitative perception of the three-dimensional environment around the vehicle, just like people without eyes, the self-driving decision-making system can not work properly.

In order to perceive safely and accurately, the autopilot system uses a variety of sensors, among which ultrasonic radar, millimeter wave radar, lidar, and cameras can be broadly classified as "vision".

Compared with other application scenarios, the application scenario of autopilot is relatively complex, especially in the complex and changeable road environment, the breakthrough of perceptual technology can not be solved by the innovation of algorithm or technology.

In this case, the manually tagged data with rich semantic information can make the algorithm better understand and identify the picture information and obstacle information transmitted by visual cameras, lidar, millimeter wave radar and other sensors, and comprehensively improve the perception ability of self-driving vehicles.

Autopilot tagging scene (source: Manfu Technology)

However, not all the tagged data can be used by self-driving cars. Theoretically, the more accurate the labeling results of the data, the better the results of the algorithm. Therefore, large-scale and accurate tagged data sets will have a substantial boost to the commercialization process of autopilot.

Baidu also talked about this when it opened up ApolloScape:

"although there are many data sets to choose from abroad, the complexity of road conditions at home is obviously different from that abroad. Although we can often hear some companies announce the public testing of their self-driving cars, most of the tests are conducted under relatively simple road conditions, and under complex road conditions, self-driving is far from reaching the road standard, many of these problems are difficult to solve only by technical means and need the help of large-scale accurate data sets, which is the original intention of Baidu's decision to open ApolloScape. Baidu hopes to attract more enterprises and developers to use and supplement data sets in an open way, so as to improve the perception of autopilot. "

From the above point of view, high-quality tagged data sets have become one of the key factors to assist the large-scale commercial landing application of self-driving vehicles.

However, the current data tagging industry still lacks the ability to provide high-quality annotated datasets.

Relevant data show that the current data tagging industry has a single delivery rate of less than 50%, and a delivery rate of less than 90% within three times, which is far from meeting the needs of AI enterprises.

On the one hand, it is related to the lack of high-quality data annotation tools, on the other hand, it is also related to the industry's over-reliance on manpower. The problems existing in the data tagging industry have affected the commercial application process of many AI projects to a great extent.

From a long-term point of view, as the combination of artificial intelligence and various industries becomes more and more close, refined, scene-oriented and more dimensional data become more and more important to the industrialization of AI, the need for change in the data tagging industry is imminent, and high-quality data is the real future of the industry.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.