Learning Cascaded Reduced-Set SVMs Using Linear Programming

This paper proposes a simple and efficient detection framework that uses reduced-set kernels. We first describe our approach which reduces the number of kernels. A convex optimization method is used for calculating the reduced sets. Following this, we propose a method that optimally designs the cascade. Our experimental results indicate that our method minimizes complexity regarding the number of kernels in the cascaded structure while preserving the low error rates. Our algorithm generates the optimal weight of kernels for each cascade stage. This proposed algorithm achieves high detection-rates at low computational cost.