What Do We Need to Train Self-Driving Cars Faster? More Data
Tom Spendlove posted on February 15, 2018 |
Scale API has released a new set of tools to collect more data and train self-driving vehicles.

The engineers at Scale API say that the release of autonomous vehicles has been less than breathtaking. Instead of fleets of free range vehicles roaming the streets we’re seeing small scale adaptions in localized areas. The company’s solution for faster adaption of self-driving vehicles is data. They believe that the more data available to vehicle manufacturers, the faster we can adapt to autonomous vehicles. Today Scale launched Sensor Fusion Annotation, an application programming interface to "take the data ingested by autonomous cars, like un-labeled datasets, to Scale, who uses AI and human intelligence to annotate it for the training of computer vision models."

Several different annotations are available as data and demonstrated on the company’s website. LIDAR and RADAR identify objects in point clouds and draw cubes around them, telling the user position and size of the items. Semantic segmentation can classify an image’s pixels and give information about the image on a pixel by pixel basis. Polygon annotation creates boundary polygons around user-instructed objects and tells the user the vertex positions of the polygons. Line annotation shows the lane markers in a road. Point annotations can tell a user the location of specific objects that are of interest, and the user’s distance from those objects. Cuboid annotation shows objects and builds 3D cuboids around the object in camera images. The documentation page contains programming chunks that demonstrates how many of the functions are carried out using JavaScript, Python, Ruby and cURL. The annotation APIs are proprietary so there are quite a few demonstrative chunks of code on the site but no batches of data. The self-driving car page shows several of the functions as demonstrations allowing a user to “run the code” and see what happens between a photograph and the data return to the user.

One surprising fact for me is the level of expected adaptation for autonomous vehicles. Last year the statistic I kept hearing was 10 million cars on the road by 2020 but the current figures say 238 million autonomous vehicles by 2030. As we get closer and closer to our expected autonomous vehicle future it will be interesting to see how manufacturers gather and use data to make for better and safer user experiences.

(Images courtesy  Scale API)

Recommended For You