Optimal Cascade Construction for Detection using 3D Models

Hon-Keat Pong, Tat-Jen Cham

Abstract: We describe a method for optimal construction of a detection cascade comprising 3D models of increasing levelof- detail (LOD). An LOD 3D model hierarchy of the target object is first generated. By analyzing detection performance of each individual model in the LOD hierarchy, an optimization framework that allows trade-off between speed and accuracy is formulated. The formulation allows models to be explicitly selected for inclusion in the final detection cascade while achieving optimal running time with respect to a target detection performance.