@Ingemar:
isn't so easy to do on the GPU.
And that would be the challenge in this method. And it is cheating in the sense that it is in the physics and advanced math forum and my implementation shamefully uses Prometheus code in areas where I am too lazy to go back into OpenGl (although it could just use gl)...

@Phibermon:
Or if you broke up the collision detection into segmens and scanned those for even a single pixel you could use that to get 'pixel perfect' to the nearest group of pixels depending on hardware

And with a velocity variable, you'd be all set to go as to where the collision is from with some simple math from the near centre of the colision pixel group - of course that would use up a lot of cycles but would yield an accurate result.