Non-parametric CRFs for Image Labeling

Jeremy Jancsary, Sebastian Nowozin and Carsten Rother
NIPS Workshop on Modern Nonparametric Methods in Machine Learning
December 2012, Lake Tahoe, Nevada


We introduce a powerful non-parametric image labeling framework, Regression Tree Fields (RTFs), and discuss its application to image restoration. The conditional structure and the parameters of our model are estimated from training data so as to directly optimize for popular performance measures, resulting in excellent predictive performance at low computational cost.

Main article