Contents
Yandex developers have released the world’s largest dataset for teaching unmanned vehicles to predict on the road
What’s going on
- Yandex has developed and published a dataset for training algorithms for predicting the movement of vehicles and pedestrians. It includes data equivalent to 69 hours of continuous driving – this is the largest such dataset in the world.
- It consists of 600 scenes of 10 seconds each, each of which is a map of the area with markings, mapped cars, pedestrians and traffic lights, as well as indicating the parameters of moving objects and weather conditions.
- The developers pursued the goal of preparing the neural network for a data distribution shift, when the model gets into situations unfamiliar to it, because it has not seen such data and does not know how to behave. That is why Yandex collected information in countries with different weather conditions and driving styles – our country, Israel and the United States.
- In addition, the Yandex team wants to teach the algorithm to evaluate the uncertainty about its decisions, thereby increasing its efficiency.
- Together with colleagues from Cambridge and Oxford universities, Yandex announced the Shifts Challenge competition to create algorithms for predicting the movement of cars and pedestrians. The essence of the task is as follows: based on the first five seconds from the dataset scene, predict the movement of all objects in the next five seconds, and also estimate the uncertainty of the forecast.
What does it mean
Logic algorithms used in unmanned vehicles are currently not as developed as environmental perception algorithms. At the same time, it is they who are responsible for predicting the situation and helping to pave a safe path for movement. The development of neural networks responsible for the “logic” of the drone is not an easy task, primarily because of the so-called distribution shift.
A “shift” in driving style and some rules is something that people are able to take into account. For machine learning, this is a serious challenge: to train the algorithm so that in a situation that is far from the training sample, to issue a not entirely delusional forecast, to correctly take into account the shift relative to the training sample. For example, if the model was taught on a sample where everyone passes people moving in a circle in a roundabout, it should not provoke an accident or fall into a stupor in a country where this rule is ignored, ”Andrei Sebrant, director of strategic marketing for Yandex services, commented on the news. your telegram channel.
It is worth noting that the number of locations for testing the drones of most developers is very limited. Thus, when faced with completely new driving conditions, the probability of making the wrong decision by the algorithm due to the distribution shift increases. That is why it is so important to train the neural network to apply the experience already accumulated in new conditions and, on its basis, to predict the development of the traffic situation with a minimum error.