Question

Say I wanted to downsample an image in realtime (1280×720) to a very small scale (16×16) and not suffer from “dancing” pixels when the image moves, which technique would I use?

This would be using XBox360/PS3 GPUs.

Note that the 16×16 image only needs to contain a very “generalised” representation of the original image – almost like a rough-estimate of colour.

Doing a direct downsample results in pixels coming in/out of the sampling so you get a dancing/disco effect when the image moves.

Thanks for any help.

Answer 1

You’ll want to use averaging instead of downsampling. For instance, setting each pixel to the average value of an 80×45 pixel block will give you a 16×16 image from a 1280×720 image. Note that that would alter the aspect ratio though…

I should clarify that you don’t necessarily have to average the entire block — you could average every other pixel in it or whatever you want. But the more pixels you account for, the more accurate the value would be. You could also use other statistical measures — the mode or median values across channels could also work.