Accelerating Computation of Exemplar-SVM by Binary Approximation based on Matrix Decomposition

The Exemplar-SVM (E-SVM) is a learning method based on exemplar that uses only one positive sample and a substantial number of negative samples. In the detection stage, it is possible to detect the location of the target object and estimate the attribute by transferring the attribute of the nearest exemplar. The use of E-SVM classifiers leads to very high computational cost because it is necessary to compute the inner products of weight vectors for multiple classifiers and an input feature vector. For accelerating the computation of E-SVM, we propose binary approximation based on matrix decomposition. First, we stack the E-SVM’s weight vectors as a matrix. Then, we decompose the matrix into common binary basis vectors and real-valued coefficient vectors for computing the approximated inner products by logical operation. We also introduce early rejection by cascade structure classifier into the proposed method. The evaluation experiments show that the computation time of the proposed method is lower by a factor of 200 than that of the conventional E-SVM.