Taking two dimensions from the autoencoder is just a specific example. You can get any pair of dimensions and plot the results. In each case, it is common to see that the separation of the three classes is not good enough when compared to the PCA output.