Learning Low-Level Vision

model image dependencies using a Markov network
use loopy belief propagation
for compression and generalization, use PCA to find lower dimensional
 representation for patches.  measure distances in this space using
 Euclidean metric
use an instance-based output: the estimated scene in each patch is
 always some example from the training set ("line up suspects").
interesting combination of dimension reduction with something like
 constraint satisfaction to find a coherent interpretation for the
 input data.
for motion example, used tree-structured VQ method to reduce
 dimensionality of image patches to one in 300.
"For this problem with this training data, the machinery leads to
 figure/ground segmentation, aperture problem constraint propagation,
 and filling in".  Exciting!
The method requires supervised data (knowing the image *and* the
 underlying scene).  Hard to know how to apply this to an autonomous
 system.

Image Analogies

Notes the key role of similarity.
Oh, also uses Markov random fields.
Fails on long-range coherent features (brush strokes).  (Perhaps the
 techniques from the other paper could be used.)
"continues in the two-word-title tradition of these earlier
  works". :-)
uses a feature-vector representation of pixels (RGB + luminence,
 etc.).
Analogy is A:A' :: B:B'.  Given A, A', B, generate B'.
Finds a pixel in B that is a good match for a pixel in A.
Uses two kinds of similarity: approximate search (content based),
 coherence search (context based).  Kappa trades these off
 (an application dependent parameter).
uses approximate nearest neighbor match and some PCA for dimension
 reduction in high dimensional feature space examples.
Did matching in luminance space when possible.
Includes various hacks to speed things up.