Computing the three-dimensional structure of scenes from multiple images requires the same scene points to be identified in those images. This is the correspondence problem and is a fundamental problem to most computer vision applications. Correspondence errors are often caused by point mismatches (called outliers) and are typically non-Gaussian in nature. This means that standard least squares methods will not work very well. Robust methods are able to handle outliers and RANSAC (and its variants) is the go-to method for structure computation and as far as I know is universally used. In this talk I will show that an alternative to RANSAC, called case deletion, is just as accurate and often significantly faster than RANSAC, and is more appropriate for real time applications.
Last modified: Tuesday, 09-Sep-2014 09:07:17 NZST
This page is maintained by the seminar list administrator.