AVT Expo Interview: Why manually annotated data no longer cuts it

LinkedIn +

Heikki Laine, VP of product and marketing at Cognata, explains why manually tagged data is no longer sufficient for the effective training of AVs and suggests a solution. He will talk more about this in his presentation at the Autonomous Vehicle Test & Development Symposium. For more information, up-to-date programs and rates, click here.

What is the background to your presentation?

Training and testing the AI in the computers that operate automated driving systems takes huge amounts of data. Historically, all of this data is collected during physical testing, but in order to make the collected data useful, it must be accurately and extensively labeled through annotation.

This process is extremely human labor intensive. It can take hundreds of hours to annotate just a single hour of driving data. The end result is a corpus of training data that is not only costly to create, but also limited in size and variety, because it contains only those locations, scenarios, and sensor configurations encountered or used by the test vehicle. To make matters worse, the data can be inconsistently or inaccurately annotated due to the human element of the process. In short, the gaps in variety, accuracy, and scale of annotated data are three of the biggest challenges facing system developers today.

How are you solving these issues?

Heikki Laine CognataAt Cognata, we are leveraging our own deep experience in AI to bridge this gap. As a provider of simulation software, Cognata has developed a novel method of generating highly photo-realistic environments using deep learning. Our purpose-built deep neural networks are trained on actual automotive-grade sensor data and allow us to project virtual environments as they would be perceived by real sensors. We are applying this same technology to the creation of datasets for training and testing.

Datasets created in virtual environments have a number of compelling advantages. They can include a near-endless variety of weather, road surface and lane marking conditions, as well as rare or dangerous elements. The datasets can be re-created if a sensor position is changed or a new sensor is added and – perhaps most importantly – the annotations are accurate and consistent because they are machine-generated. More accurate data reduces the number of training iterations necessary for vehicle AI.

What do you see as the ideal balance and relationship between simulation and real-world testing for AV and ADAS?

Physical testing will continue to play a very important role in the training and testing of automated driving systems for both AV and ADAS applications, alongside advanced simulation. For as long as vehicles will be sharing the roads with humans, people will want to test and evaluate vehicle behavior in the real world. And rightly so. In the meantime, we will continue to see simulation accelerate the pace of development and increase safety by focusing and refining the use of physical testing.


Catch Heikki’s presentation titled “Virtual validation and simulation at large scale, for training, testing and deploying automated driving systems” at the Autonomous Vehicle Test & Development Symposium. For more information, up-to-date programs and rates, click here.

Share this story:

About Author


Comments are closed.