Benefits of Separable, Multilinear Discriminant Classification

This paper presents an empirical investigation of the merits of tensor-based discriminant classification for visual object detection. First, we briefly discuss 2D separable discriminant analysis for grey value image analysis. Then, we contrast this tensorial approach with classical linear discriminant analysis. Our findings on a standard data set for object detection in natural environments show that, for the task of image analysis, tensor-based discriminant classifiers perform very robust. They learn and run faster and also generalize better than conventional techniques based on vectorial representations of the data.