Title:StructBoost: Boosting Methods for Predicting Structured Output Variables

Abstract:Structured learning has found many applications in computer vision recently. Analogues to structured support vector machines, here we propose boosting algorithms for predicting multivariate or structured outputs, which is referred to as StructBoost. As SVM-Struct generalizes SVM, our StructBoost generalizes standard boosting such as AdaBoost, or LPBoost to structured learning. AdaBoost, LPBoost and many other conventional boosting methods arise as special cases of StructBoost. The resulting optimization problem of StructBoost is more challenging than SVM-Struct in the sense that the problem of StructBoost can involve exponentially many variables and constraints. In contrast, for SVM-Struct one usually has an exponential number of constraints and a cutting-plane method is used. In order to efficiently solve StructBoost, we propose an equivalent 1-slack formulation and solve it using a combination of cutting planes and column generation.