This paper investigates the performance of Support Vector Machines with linear, quadratic and cubic kernels in the problem of recognising 3D objects from 2D views. It describes an experiment using the complete set of images from the Columbia Coil100 image database. Image views were randomly selected from the object classes. Previous works used only subsets of the classes, from which only a few training and testing set sizes were extracted and object views were usually too close to each other, which may have artificially increased the recognition rates. In our experiments, we observed that the degree of the polynomial kernel played a minor role in the final results. Moreover, although recognition rates were slightly inferior to those of previous work, a clearer picture of the SVM performance on the Coil100 image database has been produced.