Higher-order clique based image segmentation using evolutionary game theory

Abstract

This paper describes a novel algorithm for labeling problems of image segmentation. Beyond the pairwise model, our proposedmethod enables exploration on cliques, which are able to capture rich information of the scene. However, the dilemma isthat, while our objective is to assign each pixel a label, the cliques are only limited to work on sets of pixels. To address thisproblem, the interaction between pixel and clique is studied. The labeling problem is solved using iterative scheme incorporatingExpectation-Maximization (EM) algorithm that: in the E step, we would like to estimate labeling preference of pixels fromclique potentials with known labeling distribution; and then update clique probabilities in the M step. We optimize the proposedfunction in the framework of evolutionary game theory, where the Public Goods game (PGG) is employed. Taking the advantageof large size cliques, our algorithm is able to solve multi-label segmentation problem with effective and efficiency. Quantitativeevaluation and qualitative results show that our method outperforms the state-of-art. Especially, we apply the proposed algorithmon urban scene segmentation, which aims at segmenting geometric inconsistent objects via vertical assumption. We believe thatour algorithm can extend to many other labeling problems.