I have experimented on using KPCA as a pre-processing method for some classification problems with both low and high dimensional input spaces. In general the choice of kernel and especially the number of eigenvectors (more is not always better) used for classification has a pretty significant effect on classifier performance. The good news is that using KPCA with a linear classifier allows one to look at the projections of the data on the KPCA eigenvectors directly to understand why the classifier does well or badly on a particular dataset (i.e. look for linear separability in the feature space). The bad news is that in most cases using KPCA doesn't improve classifier performance over just using normal PCA. In any case the method is so easy to implement and fast (at least for small data sets) that it is worth experimenting with especially since even a simple LDA classifier can have very good performance when KPCA is used as a pre-processing step.