Learning generative visual models from few training examples: An incremental Bayesian approach tested on 101 object categories