New article on my site: HyAB k-means for color quantization

In which I try to improve color clustering with a different distance function. It's a simple technique in the end but it's pretty hard to evaluate if it's an improvement or not!

https://30fps.net/pages/hyab-kmeans/
https://github.com/pekkavaa/HyAB-kmeans/ with #numpy code

This documentation page on #Blender's internal mesh data structure is really good: https://developer.blender.org/docs/features/objects/mesh/bmesh/ It has very thoughtful comparisons to half-edges.

I stumbled upon it when researching mesh libraries for Python that support N-Gons. Perhaps I'll try reimplementing BMesh myself.

Measured a 256-shade greyscale ramp through my capture card with both on my laptop and the Nintendo 64 (via Retrotink). The response is linear in both (phew!) but I had to mess with the capture card brightness & contrast settings to make the N64 closer to the input. Still clips the whites too early.

#n64 #n64dev

Today I was scratching my head why k-means didn't seem to reduce Mean Squared Error. The clustering seemed fine. Is the error computation broken? How can it be when it's so simple:

def compute_mse(a, b):
return np.mean((a-b)**2)

Well, arguments 'a' and 'b' were images with 8 bits per channel, so...

#numpy

Another example on how mean squared error (MSE) doesn't predict final image quality. I implemented Variance Cut Color Quantization in two ways, left is with greedy split plane optimization and on the right side it's with a few rounds of global k-means that results in lower MSE.

I actually prefer the image on the left side since its more detailed.