Write a road map

[apt1002]

Requested by @rrthomas and a good idea.

[apt1002]

This is as good a place as any. In no particular order, here are some things I would like to do eventually:

• [ ] Generalise beyond greyscale images:

  • [ ] Add an axis to every image transform to represent the colour channel index (#1).
  • [ ] Convert to an internal colour space similar to LUV.
  • [ ] Maybe predict U and V based on L.
  • [ ] Generalise the perceptual model to estimate colour discrimination ellipsoids.
  • [ ] Generalise the quantizer. Probably just loop over the colour channels.
  • [ ] Same tasks for the alpha channel.

• [ ] Work on the tree quantizer:

  • [ ] Abstract over the perceptual model, to get it out of the way and prevent bugs.
  • [ ] Choose a codebook (a finite set of common nearly blank image tiles), and make each leaf of the quantized tree be one of these tiles.
  • [ ] Write Tree::arrow() analogous to ShiftedBCC::arrow():
  • [ ] Write an arrow() function for codewords.
  • [ ] Define how non-leaf trees map onto leaf trees.
  • [ ] Work out the nearest-neighbour algorithm.
  • [ ] Define a probability model over trees.
  • [ ] Modify the nearest-neighbour algorithm to balance quantization error with file size.
  • [ ] Write an iterative algorithm to train the codebook on a corpus of images.

• [ ] Write an arithmetic encoder (probably not worth adding a third-party dependency):

  • [ ] Interface it with the tree quantizer.

• [ ] Define the FVQ file format.

  • [ ] Header:
  • [ ] Version
  • [ ] has_colour
  • [ ] has_alpha
  • [ ] Original resolution
  • [ ] Quality
  • [ ] Container format.

• [ ] Optimization:

  • [ ] SIMD
  • [ ] Threads
  • [ ] Hand-compile some of the overly abstract parts.