| |
Abstract:
A goal of low-level vision algorithms is to infer underlying
scene information (such as velocities, shapes, or reflectances)
from an image or set of images. We introduce a simple,
training-based method: learn the probability of every scene
interpretation for any local image patch, and the probability that
any local scene neighbors another. The former probabilities give
scene estimates from local image data, and the latter allow the
local estimates to propagate. We use Markov networks to optimally
propagate the local information, represented as mixtures of
gaussians. We demonstrate the technique for two problems: motion
estimation, and the disambiguation of shading and reflectance. We
find that the local statistical information, without added problem
domain knowledge, finds good global scene interpretations.
|
