Always pay attention to open and continue to suck to absorb new members to join the Baidu Apollo automatic driving platform, "recently uncharacteristically" initiative announced to join the California University Berkeley DeepDrive automatic driving industry alliance.
In the US time March 8th, Baidu announced that the Apollo automatic driving platform was formally joined in the DeepDrive deep learning automatic driving industry alliance, and released the Apollo automatic driving data set ApolloScape.
To attract Baidu Apollo platform to actively join the industrial alliance, I am afraid the latter is a more abundant automatic driving academic achievement and industrial resources.
If you know enough about UC Berkeley (University of California at Berkeley), you know that DeepDrive is one of the two UC Berkeley related laboratories for automotive intelligence (the other is InterACT).
The results of DeepDrive does not stay in the lab, but closely integrated with the industry, the current partners such as BOSCH, ZF and other first tier suppliers, Volkswagen, Honda, Hyundai car prices, NXP, NVIDIA, HUAWEI and other chip makers, and controlling potential companies are China its partner.
Partners in the *Deep Drive research project
The DeepDrive deep learning automatic driving industry alliance is an industrial alliance led by University of California at Berkeley, which is applied in the field of computer vision and machine learning in the automotive field.
Its members include: NVIDIA, Qualcomm, Ford, GM and other 20 world top enterprises in the field of automatic driving, the research project covers the perception, planning and decision-making, and other key areas of automatic driving deep learning.
The purpose of Baidu's entry into the alliance is to share the research results and accelerate the technology innovation and landing application process of automatic driving by working together with the leading companies and the top academic research institutions in the world, so as to enhance the R & D strength of the automatic driving.
ApolloScape: the amount of data is more than 10 times more than the same data set
Another view of the release is Baidu's open ApolloScape data set.
Data sets are generally divided into two categories: one is a common data set, a data set proposed by a purely computer vision domain, and this data set is based solely on having
Obviously, Baidu also wants to build ApolloScape into a data set like this. So, what are the highlights of the ApolloScape dataset?
Baidu believes that the massive real data, high quality is essential in the development of automatic driving test "raw material", so the ApolloScape data is similar data sets (such as Cityscapes) more than 10 times.
Among them, the content of data includes the high resolution image data of several hundred thousand frames by pixel semantic segmentation, such as perception, simulation scene, road network data and so on. In Baidu, the ApolloScape data set covers more complex road conditions from the data difficulty dimension. One example is a single image of as many as 162 vehicles or 80 pedestrians.
In addition, the open data set uses pixel by pixel semantic segmentation and labeling. Baidu calls it the most complex and accurate data and the most automated data set. "
ApolloScape annotation data example
ApolloScape depth data example
Comparison of data instances by Kitti, CityScapes and ApolloScape
Another feature of ApolloScape is the high resolution image data that contains hundreds of thousands of frames per pixel semanteme.
For the use of data is convenient for researchers to better set the value of Baidu defined a total of 26 different meanings of language data instances in the data (such as automobile, bicycle, pedestrian street, buildings, etc.), and will further cover the more complex environment, weather and traffic conditions.
Information of the various instances contained in the data
Simulation is also a key project of this data set. The goal of Baidu is to create a simulation platform with the highest reduction and the most abundant scene in the real world.
According to Lei Feng network understanding, based on the Apollo simulation platform, the ApolloScape program will automatically drive the vehicle into dozens of vehicles traveling the same road network, via the simulation of complex and multi vehicle driving scene game process, to help the developers to effectively test and optimize the prediction, decision and path planning algorithm, improve the automatic driving diversity test.
In order to revitalize this data set, to attract more developers to use the ApolloScape data set, during this year's CVPR, Baidu Apollo will jointly held at University of California at Berkeley (Workshop on Autonomous automatic driving workshop Driving), hope to provide a breakthrough in technology innovation and application platform for the development of automatic driving to global and researchers.
"Large systems" and "small modules"
In the past, one of the common problems that computer vision has been faced with is that the old algorithms do not work on the new dataset.
"We claim to solve a problem is just to solve a data set, and it doesn't mean that we really solve this problem, which happens frequently. "CTO, a domestic autopilot, said to Lei Feng (public number: Lei Feng net).
For example, we can dismantle the "large system" of autopilot into 100 small computer visual problems.
But there are two points worth pondering. First, we don't know which question is more important in the 100 questions. Second, we don't know which problem and the extent to which we can solve the problem of automatic driving.
So how to solve the problem between automatic driving "big system" and automatic driving "small module" is the next advantage that Baidu ApolloScape data set needs to establish. It is also an automatic driving data set that practitioners and developers need.