Abstract
We introduce an algorithm,SVM-IS,for structured SVM learning that is computationally scalable to very large datasers and complex structural representations.We show that strucrured learning is at least as fast-and often much faster-than methods based on binary classification for problems such as deformable part models,object detec-rion,and multiclass classification,while achieving accura-cies that are at least as good.Our method allows problem-specific structural knowledge to be exploited for faster op-rimization by integrating with a user-defined importance sampling funcrion.We demonstrate fast train times on two challenging largescale datasets for two very different prob-lems:ImageNet for multiclass classification and CUB-200-2011 for defonnable part model training.Our method is shown to be 10-50 times faster than SVMSC for cost-sensirive multiclass classification while being about as fast as the fastest 1-vs-all methods for multiclass classification.For deformable part model training,it is shown to be 50-1000 times faster than methods based on SVMStruCt,min-ing hard negatives,and Pegasos-style stochastic gradient descent.Source code of our method is publicly available.