A question - statistically how much better is it than producing OBBs via applying KLT/PCA on the (x,y,z) coordinates, obtaining major/minor covariance axis? I used that one for my SW 3D engine and was quite happy about it - very fast eigenvector computation and only orthogonal projections needed.
I don't have any numbers, but I think they mentioned this in the article: It depends on how well distributed your points are. There are a lot of cases where a point cloud will yield an inefficient obbox using the PCA method, while (if what you want is the minimum volume) their method guarantees optimality. As always, it depends on your use case.
Sure, that issue was prevalent with low-poly models but I doubt it's an issue with current model complexity. I basically wanted to understand how much of an improvement can I reasonably expect shall I decide to implement their algorithm.
