Tensor Discriminant Analysis for View-based Object Recognition

In this paper, we use a general Mth order tensor discriminant analysis approach (Tao et al. 2005) for view based object recognition. This method is an extension of the 2D image coding technique (Shashua and Levin, 2001) to general Mth order tensors for discriminant analysis, and has good convergence property. We demonstrate the performance advantages of this approach over existing techniques using experiments on the COIL-100 and the ETH-80 datasets. Specifically, our experimental results on ETH-80 show the particular strength of this tensor discriminant analysis method when only a small number of training samples with big intra-class variation are available