Fast SVM Approximations for Object Detection
It has been shown that Support Vector Machines (SVMs) provide state-of-the-art accuracies in object detection. However, for real time applications, standard SVMs are usually too slow. In this work, we propose a method for approximating an SVM detector in terms of a small number of separable nonlinear filters. We are building on work of Romdhani et al. (ICCV 2001), where an SVM face detector was approximated using the so-called reduced set algorithm and evaluated in a cascade. However, when using plain gray values as features, we found it more effective to reduce the high computational cost for the pixel-wise comparisons, rather than focusing on sparsity of the detectors alone. In our approach, we constrain the reduced set optimization to a class of nonlinear convolution filters which can be evaluated more efficiently (i.e. O(w+h) instead of O(wh), where w and h are the patch dimensions, respectively). We demonstrate a prototype of our system which runs in real time on a standard PC.