 This is neat and the illustrations are great. A few things that will probably give large gains while being low hanging fruit:1. There are triangle interpolation schemes out there now that are smoother than barycentric coordinates which should give much better results.2. Look up DDE - Data dependent triangulation. It switches edges to connect points to neighbors that have similar values. It will get rid of some of the spikiness and leave more smooth gradients.3. The running the edge detection twice scheme mentioned in the comments works because you want the change of the gradient, and you need both sides represented. So the double edge detection will give you manifolds, which is good.4. Instead of having arbitrary vertex positions, you can just specify the offset to the next point. Then instead of an x and y value you can use one (possibly uint8_t) value to encode where the next point will go.5. You can also chop some accuracy off of colors. In RGB, you can lose accuracy in blue and some in red. In other schemes like you can keep accuracy in luminance and lose it heavily in hue and chroma/saturation, etc. W.r.t. running the edge detection twice. Is there a name for this operation? Finding points that are near an edge but not on the edge? The Laplacian. Weird. I hadn't made the connection with Physics there. I suppose it is a general mathematical operator.Thanks! Yes, it's not exactly equivalent here but it's pretty close. Basically, Sobel kernels are analogous to the 1st derivative and the Laplacian is analogous to the 2nd. You'll also often see the Laplacian combined with a Gaussian (the "LoG" operator) for pre-smoothing since it tends to be particularly sensitive to any noise. Being a second derivative, I suppose it would be.So does that mean that there are other kernels that approximate derivatives "better"? Like with finite differences? Applications are open for YC Summer 2019

Search: