First page Back Continue Last page Overview Text


Now before switching to the technical part of my talk, I’ll briefly sketch and compare common approaches to the stereo correspondence problem. First is a so called local approach. In LA, we look at one image patch (or window) at a time, that is we enforce the 2 constraints that we have on a small image patch. Thus we are solving many small problems independently. It’s fast but generally not the most accurate.
In the global approach, we look at the whole image simultaneously, that is we enforce the 2 constraints on all pixels at the same time. Thus we solve one large problem. It’s usually slower than the local approach, but generally more accurate.
These 2 approaches have been around for a while, but we are going to make significant improvements to both of them using graph algorithms.