Enumerate! Don't Estimate.

Shai Avidan (Tel Aviv U., Isreael)


However estimation is difficult because of noise in the measurements, missing data and outliers. On top of that, it is not always possible to give global guarantees for the estimated solution. Enumeration, on the other hand, is simple. It requires going through all possible hypotheses, evaluating them one by one and reporting the best one. The key is to prove that there is an efficient way to enumerate all possible hypotheses. I will show a couple of applications that find a guranteed approximation to the global optimum for a number of computer vision problems including 2D affine template matching and 3D symmetry detection. Joint work with Simon Korman, Roee Litman, Daniel Reichman, Gilad Tsur and Alex Bronstein.