Soft Histograms for Belief Propagation

Belief propagation methods are powerful tools for various problems in computer vision. While most methods assume a discrete set of labels for each node in the graphical model, there has recently been an increased interest in using real-valued labels and continuous probability density functions for such problems. We propose using channel representations (soft histograms) as a new way of moving from discrete to real-valued labels. The soft histograms are related to continuous density functions through the maximum entropy principle, and a complete soft histogram-based belief propagation method is evaluated and compared to hard discretization methods on simulated and real data.