Accelerated Training of Linear Object Detectors

We describe a general and exact method to speed up the training of linear object detection systems operating in a sliding, multi-scale window fashion, such as deformable part-based models. Our approach consists of reformulating the computation of the gradient as a convolution, and making use of properties of the Fourier transform to obtain a speedup factor proportional to the linear filters' sizes. This technique does not rely on the sparsity induced by a specific loss, nor on a stochastic sub-sampling of the training examples. Experiments on the PASCAL VOC benchmark show a speedup factor of more than one order of magnitude compared to a standard exact generic method.