Jigsaw Puzzle Solving as a Consistent Labeling Problem

Abstract

We explore the idea of abstracting the jigsaw puzzle problem as a consistent labeling problem, a classical concept introduced in the 1980s by Hummel and Zucker for which a solid theory and powerful algorithms are available. The problem amounts to maximizing a well-known quadratic function over a probability space which we solve using standard relaxation labeling algorithms endowed with matrix balancing mechanisms to enforce one-to-one correspondence constraints. Preliminary experimental results on publicly available datasets demonstrate the feasibility of the proposed approach.

Publication
In International Conference on Computer Analysis and Patterns
Alessandro Torcinovich
Alessandro Torcinovich
Postdoc in Machine Learning

My research interests include Game Theoretic, Deep and Weakly Supervised Learning