ECCV 2018 Daily - Tuesday

Daily Tuesday 11 11 Lyft Holger Rapp , a software engineer at Lyft Level 5 , tells us that they face a number of computer vision challenges. One of the biggest is calibration , which needs to be constantly re-evaluated, reconsidered, and improved. For example, on the self-driving cars they have very tight control over the calibration of the cameras and the sensors, but at Lyft they have a large fleet of human- operated cars and would like to make more use of them than they currently do. One of the challenges of making use of this data is figuring out how the cameras in those cars are oriented. If there are two cameras in there, what is the stereo distance? What is the calibration of that? How can they actually get useful data that they can directly apply to self-driving cars out of this massive fleet? Holger explains how they might solve this: “ We have many, many approaches towards solving this. There are very classical approaches that we pre-calibrate and put the cameras in after. That’s a lot of manual labour. We try auto- calibration routines. We try playing around with different sensors and augmenting with different sensors. Of course, we’re also looking at novel ideas. For example, using machine learning to segment images out and then using known sizes of objects that we have classified in those images to calibrate. But it remains a tough problem. For this particular approach, there aren’t really any semi- autonomous solutions. For the self- driving cars, we can do everything fully automated. For the large fleet of human-driven cars, there’s just too much work. There’s nothing that human operators can really help with .”