An Adversarial Optimization Approach to Efficient Outlier Removal

摘要: This paper proposes a novel adversarial optimization approach to efficient outlier removal in computer vision. We characterize the problem as game that involves two players of conflicting interests, namely, model optimizer and outliers. Such an view not only brings new insights into some existing methods, but also gives rise general framework provably unifies them. Under proposed framework, we develop is able offer much needed control over trade-off between reliability speed, which usually available previous methods. Underlying mixed-integer minmax (convex-concave) formulation. Although generally amenable optimization, show for commonly used vision objective functions, equivalent Linear Program reformulation exists. significantly simplifies optimization. demonstrate our method on representative multiview geometry problems. Experiments real image data illustrate superior practical performance recent techniques.