Taxonomy of Dual Block-Coordinate Ascent Methods for Discrete Energy Minimization

TL;DR

This paper provides a comprehensive taxonomy of dual block-coordinate ascent methods for discrete energy minimization, unifies existing algorithms under a single framework, and introduces a new state-of-the-art solver with improved performance.

Contribution

It offers a unified framework for understanding dual block-coordinate ascent solvers and proposes enhancements leading to a superior solver for discrete energy minimization.

Findings

Existing solvers have sub-optimal updates that can be improved.

The new solver outperforms previous methods across various problem instances.

The framework enables systematic exploration of solver variants.

Abstract

We consider the maximum-a-posteriori inference problem in discrete graphical models and study solvers based on the dual block-coordinate ascent rule. We map all existing solvers in a single framework, allowing for a better understanding of their design principles. We theoretically show that some block-optimizing updates are sub-optimal and how to strictly improve them. On a wide range of problem instances of varying graph connectivity, we study the performance of existing solvers as well as new variants that can be obtained within the framework. As a result of this exploration we build a new state-of-the art solver, performing uniformly better on the whole range of test instances.