Middlebury College
Department of Computer Science Seminar
Ground Truth for Computer Vision:
How to get it and what to do with it.
Daniel Scharstein
Department of Computer Science
Middlebury College
What is "Ground Truth"? The term originates in remote sensing: for example, elevation maps generated from aerial photographs must be related to the true elevation and location of features as measured on the ground. In computer vision, we need to know the true depth or true object motion if we want to evaluate the performance of stereo vision or optical flow algorithms.
In this talk I will show different techniques for creating image datasets with ground-truth geometry, including structured lighting, laser and CT scanners, and – as of this month! – fluorescent and glow-in-the-dark paint. Much of this work was done in collaboration with student researchers in our robot and vision lab here at Middlebury.
I will then focus on applications of data with ground truth. An obvious one is testing – in fact, the Middlebury datasets are now the standard stereo vision benchmark, used by researchers around the globe. But data with ground truth can also aid in developing new algorithms, and I will briefly describe two recent research projects: designing robust stereo matching costs, and automatic learning of parameters in a conditional random field stereo model.
The talk will be accessible to a general audience.
Friday, March 16, 2007
12:30 p.m. to 1:15 p.m.
McCardell Bicentennial Hall 538
Lunch will be provided at 12:15 p.m.
All are welcome to attend
This event is sponsored by the Computer Science Department.