Generic image classification has been becoming a popular topic during these years. a sparse bayesian model, namely the relevance vector machine (RVM), is combined with recently popular spatial pyramid matching kernel (SPM) to constitute an image classification system. this approach shares many advantage of other discriminative approaches, but can give a probabilistic prediction. the performance of this approach for object recognition, scene categorization and human activity classification taks are experimentally analyzed one 4 famous public datasets. the result show that the approach can achieve comparable classification rate as compared to SVM-based one, but greatly savae computational burden during prediction time. our experiments also show that it can generalize better than the SVM approach even when there is a big variation between training and testing data.