PMBP: PatchMatch Belief Propagation

PMBP is a C++ implementation of the PMBP (PatchMatch Belief Propagation) algorithm for correspondence field estimation, originally published at BMVC 2012. It provides a framework for solving stereo matching and optical flow problems by combining the efficiency of PatchMatch with the robustness of Belief Propagation using particle resampling.

Key Features

• Particle-Based Belief Propagation — Implements particle-based message passing with resampling for efficient inference
• Modular Architecture — Base classes allow easy extension for custom correspondence problems
• Stereo & Optical Flow Examples — Includes ready-to-use implementations for stereo matching and 2D optical flow
• Energy Function Customization — Provides hooks for defining custom unary and pairwise energy functions
• Particle Management — Supports particle initialization, propagation, and randomization through abstract methods

Philosophy

PMBP follows a research-oriented philosophy, providing a clean, extensible implementation of the algorithm that can be adapted to various correspondence problems beyond the included examples.