Abstract

We present a method for calibrating the Ensemble of Exlemplar SVMs model. Unlike the standard approach, which tcalibrates each SVM independently, our method optimizes htheir joint performance as an ensemble. We formulate joint gcalibration as a constrained optimization problem and deEvise an efficient optimization algorithm to find its global optimum. The algorithm dynamically discards parts of the tsolution space that cannot contain the optimum early on, Smaking the optimization computationally feasible. We exmperiment with EE-SVM trained on state-of-the-art CNN deescriptors. Results on the ILSVRC 2014 and PASCAL VOC p2007 datasets show that (i) our joint calibration procedure noutperforms independent calibration on the task of classifyiing windows as belonging to an object class or not; and (ii) wthis improved window classifier leads to better performance con the object detection task. s