Learning Generative Visual Models from Few Training Examples: An Incremental Bayesian Approach Tested on 101 Object Categories

The paper addresses learning visual object categories from only a few training images, unlike prior methods that require thousands of examples, are slow, and cannot incorporate prior information. The authors use a generative probabilistic model that represents the shape and appearance of a constellation of features belonging to an object. Prior information assembled from previously learned, unrelated categories is used in a principled way, and the model parameters are learned incrementally in a Bayesian manner. The approach is evaluated on a dataset of images spanning 101 widely varied object categories.

Experimentally, the incremental algorithm is compared to an earlier batch Bayesian algorithm and to one based on maximum likelihood. The incremental and batch Bayesian versions achieve comparable classification performance on small training sets, but incremental learning is significantly faster, making real-time learning feasible. Both Bayesian methods outperform maximum likelihood on small training sets, demonstrating the value of using prior information when learning from few examples.