Improving Recognition through Object Sub-categorization

We propose a method to improve the recognition rate of Bayesian classifiers by splitting the training data and using separate classifier to learn each sub-category. We use probabilistic Latent Semantic Analysis (pLSA) to split the training set automatically into sub-categories. This sub-categorization is based on the similarity of training images in terms of object's appearance or background content. In some cases, clear separation does not exist in the training set, and splitting results in worse performance. We compute the average difference between posteriors from the pLSA model, and observing this parameter, we can decide whether splitting is useful or not. This approach has been tested on eight object categories. Experimental results validate the benefit of splitting the training set.