In the Scan

Inclement Weather Not Being Considered

In late 2018, Krzysztof Czarnecki, a professor at Canada’s University of Waterloo, built a self-driving car and trained it to navigate surrounding neighborhoods with an annotated driving data set from researchers in Germany, that was not recorded in inclement weather.

From an article in Wired by Will Knight.

The vehicle worked well enough to begin with, recognizing Canadian cars and pedestrians just as well as German ones. But then Czarnecki took the autonomous car for a spin in heavy Ontarian snow. It quickly became a calamity on wheels, with the safety driver forced to grab the wheel repeatedly to avert disaster.

Some may call this an edge case – I call it reality. Shouldn’t one design your system for the worst case scenario?

The incident highlights a gap in the development of self-driving cars: maneuvering in bad weather. To address the problem, Czarnecki and Steven Waslander, a professor at the University of Toronto, compiled a data set of images from snowy and rainy Canadian roads. It includes footage of foggy camera views, blizzard conditions, and cars sliding around, captured over two winters. The individual frames are annotated so that a machine can interpret what the scene conveys. Autonomous driving systems typically use annotated images to inform algorithms that track a car’s position and plan its route. Continue reading →