Applying a convolution kernel to a regular heightmap grid to produce points of interest for a Delaunay triangulation. In other words, do edge detection on the dataset first, rather than going through the labourious process of attempting to compute errors for progressively more refined triangulations. Genius.
I’d expect the visual fidelity for a given number of points to be lower, but probably not so much lower that you couldn’t simply add a few more triangles to the rendering pipeline and get away with it.
To be investigated, I think.