ECCV 2016 Daily - Wednesday

Focal Flow - Emma Alexander, Harvard University 4 ECCV Daily : Wednesday Our Pick for Today ECCV Daily: What is the Focal Flow work you are presenting, Emma? Emma: It is a new depth queue that can be measured very efficiently. Our model is just a thin lense camera that either is on a moving platform or observes a moving scene. We show that under Gaussian Blur, and Gaussian Blur only, there’s a very simple linear constraint that relates image derivatives to a four-vector which is a similar computation to optical flow with just a 4x4 linear system over each patch. ECCV Daily: Why is this work needed? Emma: We were inspired by this new generation of micro-scale platforms that’s being developed. Computer vision has done a great job at scaling up. We’re great with big data. We’re great with big systems with a lot of power and computation. We’re starting to find out how to sacrifice computation to work on smaller systems with smaller power budgets. There are robots being built now that work with fractions of milliwatts, microwatts. Computer vision doesn’t really have anything to offer them. We found a computational shortcut for depth measurement that will allow platforms in this class. We hope to be able to understand their scenes. ECCV Daily: What is the novelty in this work? Emma: I would say the novelty is the extreme efficiency of the computation. There are a lot of ways to measure depth. We’ve got stereo, Depth from Defocus, and there are a lot of ways to handle motion. “Computer vision doesn’t have anything to offer these robots” Among the many projects presented at ECCV2016 , RSIP Vision’s engineers picked a few which they found particularly impressive, even for ECCV standards. Emma Alexander , a PhD student at Harvard University , agreed to tell us about Focal Flow: Measuring depth and velocity from defocus and differential motion . Oral Session 2B-04, today 14:00 - 15:00 Our cue is monocular. It’s passive so we don’t change the camera. We don’t shine light onto the scene. Then our measurement algorithm is just a small linear system on image derivatives. We have very few adds and multiplies. We don’t have to expend electrical power on any other part of the sensor.