Synthia virtual training ground will qualitatively change the effectiveness of learning AI robots
Alas, for the sake of justice, it is necessary to recognize that automobile-robots, which almost all leading car manufacturers are currently working on, have not yet been able to demonstrate any confident practical skills in controlling a traffic situation and are often lost, and even act contrary to any logic, increasing the likelihood Accident. But initially self-driving cars, as an alternative to a car driven by a man, were created with the aim of minimizing the percentage of accidents on the roads. Specialists from the Computer Vision Center in Barcelona, ​​who created a unique virtual testing ground for teaching machine AI in real-world situations, offered their original solution that would qualitatively improve and enhance the effectiveness of the learning process of the neural network neural network.
No one doubts that with the improvement and training of AI systems that manage robots, the latter will demonstrate more and more confident and safe driving. And to speed up the day when the concept of “backup piloting” completely lost all meaning was suggested by the team of the Computer Vision Center in Barcelona, ​​creating a virtual model of the urban infrastructure of Synthia.
The main task of the Synthia training platform, which simulates the processes of a real urban infrastructure, is to accelerate the training of artificial intelligence systems in the process of gaining the ability to recognize obstacles, instantly orient themselves and make decisions in various unforeseen situations under different weather conditions, including rain, fog, snow and ice.
')
As is known, the main material “fed” in the process of learning a neural network is a library of images and video clips taken in the real world. By analyzing images, neural networks learn to recognize, compare and classify objects of various categories, such as: other cars and vehicles, road signs, marking signs, pedestrians, etc. Using the existing database, the system interprets information from cameras in real time and sensors, taking in the final account the decision to perform braking, changing lanes, accelerating, turning, or other actions dictated by a specific traffic situation.
AI systems have no problem processing data when driving on a freeway or other straightforward trajectory. The complexity level increases many times in the city, limited visibility, in the presence of many intersections and intersections of streets, etc. On the other hand, all the images and video “fed” to neural networks should have annotation and marking of individual objects in the images, and you must manually create such annotations. In order to imagine the amount of work being done, it suffices to give an example of Daimler specialists, who, during the implementation of the CityScapes project, had to annotate nearly 20,000 images for objects, which are divided into 30 separate classes.
Mobileye, an autopilot software development company for Tesla cars, has 600 manual specialists involved in image processing, and by the end of this year their number will be increased to 1,000. The Synthia training platform (Synthetic collection of Imagery and Annotations of urban scenario), developed A team of specialists led by Herman Rosa (German Ros) from Barcelona offers an elegant solution to this uncomfortable problem.
Conceived as a video game with the participation of already classified objects, Synthia is able to simulate many situations, including the unlikely ones that the AI ​​is unlikely to face in real life.
Using Unity's game engine allows you to create a realistic model of the urban environment and spontaneously fill it with virtual life: unpredictably leading pedestrians, cyclists, correctly and incorrectly parked cars, etc. One of the program's capabilities is the ability to simulate various situations taking into account various meteorological conditions, changes time of year, etc. In a virtual environment, the researchers placed a virtual car with a fixed orientation of cameras and created for it the possibility independently navigate through the virtual city, capturing what is happening around in the pictures and videos. Thus, an artificial database was obtained from high-quality realistic images and videos with impeccable annotation. At the next stage, the data obtained can be fed to the neural network for training.
To date, the database created by experts of the Computer Vision Center has already included about 213 thousand images and videos. At the same time, all images are classified in accordance with a particular situation, with the maximum approximation imitating the real. The first field tests of the Synthia system demonstrated impressive results and high efficiency of the method.
Analyzing the learning process using eight different neural network algorithms that were fed low-quality images (with a resolution of 240,180 pixels), specialists concluded that adding “synthetic” images to the base of real images would significantly increase the recognition accuracy of objects 11 different classes from 45 percent to 55 percent. The commercial version of the software that uses high-resolution images, which is developed by the German Ros team, will increase the effectiveness of the training system even more.
To date, all data generated and accumulated by the Synthia system is freely distributed to the public under the “not for commercial use” license. This is done in order to establish feedback with stakeholders and organizations, popularize the product and find new ways to improve the platform. Synthia developers propose the creation of appropriate versions of the training platform to the creators of robobiles, taking into account the unique configuration of cameras and other virtual car sensors, which fully correspond to the configuration of cameras and sensors of the car being created.
Original source
That's all, with you there was a simple service for choosing sophisticated Dronk.Ru equipment. Do not forget to subscribe to our blog , there will be many more interesting ...
No one doubts that with the improvement and training of AI systems that manage robots, the latter will demonstrate more and more confident and safe driving. And to speed up the day when the concept of “backup piloting” completely lost all meaning was suggested by the team of the Computer Vision Center in Barcelona, ​​creating a virtual model of the urban infrastructure of Synthia.
The main task of the Synthia training platform, which simulates the processes of a real urban infrastructure, is to accelerate the training of artificial intelligence systems in the process of gaining the ability to recognize obstacles, instantly orient themselves and make decisions in various unforeseen situations under different weather conditions, including rain, fog, snow and ice.
')
As is known, the main material “fed” in the process of learning a neural network is a library of images and video clips taken in the real world. By analyzing images, neural networks learn to recognize, compare and classify objects of various categories, such as: other cars and vehicles, road signs, marking signs, pedestrians, etc. Using the existing database, the system interprets information from cameras in real time and sensors, taking in the final account the decision to perform braking, changing lanes, accelerating, turning, or other actions dictated by a specific traffic situation.
AI systems have no problem processing data when driving on a freeway or other straightforward trajectory. The complexity level increases many times in the city, limited visibility, in the presence of many intersections and intersections of streets, etc. On the other hand, all the images and video “fed” to neural networks should have annotation and marking of individual objects in the images, and you must manually create such annotations. In order to imagine the amount of work being done, it suffices to give an example of Daimler specialists, who, during the implementation of the CityScapes project, had to annotate nearly 20,000 images for objects, which are divided into 30 separate classes.
Mobileye, an autopilot software development company for Tesla cars, has 600 manual specialists involved in image processing, and by the end of this year their number will be increased to 1,000. The Synthia training platform (Synthetic collection of Imagery and Annotations of urban scenario), developed A team of specialists led by Herman Rosa (German Ros) from Barcelona offers an elegant solution to this uncomfortable problem.
Conceived as a video game with the participation of already classified objects, Synthia is able to simulate many situations, including the unlikely ones that the AI ​​is unlikely to face in real life.
Using Unity's game engine allows you to create a realistic model of the urban environment and spontaneously fill it with virtual life: unpredictably leading pedestrians, cyclists, correctly and incorrectly parked cars, etc. One of the program's capabilities is the ability to simulate various situations taking into account various meteorological conditions, changes time of year, etc. In a virtual environment, the researchers placed a virtual car with a fixed orientation of cameras and created for it the possibility independently navigate through the virtual city, capturing what is happening around in the pictures and videos. Thus, an artificial database was obtained from high-quality realistic images and videos with impeccable annotation. At the next stage, the data obtained can be fed to the neural network for training.
To date, the database created by experts of the Computer Vision Center has already included about 213 thousand images and videos. At the same time, all images are classified in accordance with a particular situation, with the maximum approximation imitating the real. The first field tests of the Synthia system demonstrated impressive results and high efficiency of the method.
Analyzing the learning process using eight different neural network algorithms that were fed low-quality images (with a resolution of 240,180 pixels), specialists concluded that adding “synthetic” images to the base of real images would significantly increase the recognition accuracy of objects 11 different classes from 45 percent to 55 percent. The commercial version of the software that uses high-resolution images, which is developed by the German Ros team, will increase the effectiveness of the training system even more.
To date, all data generated and accumulated by the Synthia system is freely distributed to the public under the “not for commercial use” license. This is done in order to establish feedback with stakeholders and organizations, popularize the product and find new ways to improve the platform. Synthia developers propose the creation of appropriate versions of the training platform to the creators of robobiles, taking into account the unique configuration of cameras and other virtual car sensors, which fully correspond to the configuration of cameras and sensors of the car being created.
Original source
That's all, with you there was a simple service for choosing sophisticated Dronk.Ru equipment. Do not forget to subscribe to our blog , there will be many more interesting ...
 |  |  |
| Why do online stores give money for purchases? | Save up to 8% on every purchase on AliExpres and other online stores in China | Return your money - choose a cashback service for Aliexpress |
Source: https://habr.com/ru/post/395257/
All Articles