Fast object and pose recognition through minimum entropy coding

We present a pattern recognizer to classify a variety of objects and their pose on a table from real world images. Learning of weights in a linear discriminant is based on estimating the relative information contributed by a set of features to the final decision. Evaluation of the discriminant is very fast, allowing for about three decisions per second on datasets without segmentation difficulties like the COIL-100 database. Experiments on that database yield high recognition rates and good generalisation over pose.