“How (Not) to Get Hit by a Self-Driving Car” is an awesome-looking interactive experience at Playable City, a collection of interactive experiences in cities around the world.
The goal of this particular experience is to traverse a 30 meter long stretch of pavement, littered with cones and other road debris, and touch the goal at the end. The big catch is that if the camera system trained on the space detects you as a human, you lose.
This is a kind of adversarial data collection, where the human participants are trying to beat the system by faking it out and slipping through without detection. The idea is to help people recognize the powers and limitations of the types of cameras that are on self-driving cars. Presumably the data would be useful to perception researchers, as well.
The video shows people turning cartwheels and draping themselves in blankets, trying to fool the camera system.
Of course, an actual self-driving car would have multiple sensors and sensor modalities (radar and lidar, typically). But this is a neat experience, just with a camera!
David Silver 