But they didn't solve the problem, Nicholas.
They propose another solution which has its own drawbacks, most notably:
- execution time
- temporal consistency (the color segmentation, and thus the result, may be very different for one time to the other, causing visual "blinking")
Having worked for 25 years in 3D computer vision, I know very well that temporal consistency can be solved by:
- either adding hacks to smooth the solution over time
- or make sure that the solution is mathematically continuous as a function of the inputs (i.e. images in this case)
Most keyers, including commercial ones and those included with Natron, implement the second solution.
Disney's solution, however, does not bring any solution to temporal consistency. The method may work, but it's pure chance. So sorry, but for me this is still work in progress, and does not deserve a 6-months work to implement a solution that probably won't work well.
My guess is that you never had to implement an image processing research paper. It is very long and difficult, and most of the time, you just can't even replicate the results of the paper on the data used in the paper.
I tried to make a result at least as good as what they show using PIK (image above), and frankly I just can't see a difference, so what's the point in implementing that other method if you have one that just works?
Let's stop the debate here. If you want it, just implement it.