Bayesian Network has been shown to be very successful for muny computer vision applications, most of which are solved using the generative approaches. We propose a novel Weighted Bayesian Network which relaxes the conditional independent assumption in traditional Bayesian Network by assigning weights to the estimations of conditional probabilities. In the Weighted Bayesian Network, the hidden variables are estimated generatively as in the traditional graphical models, and the weights of conditional probabilities are adjusted discriminatively from the training samples. The combined generative/discriminative approach in a loop preserves the advantage of generative model to perform unsupervised learning and handle missing data while improve the model flexibility and performance by the discriminative learning of probability estimation weights. Our experiments show a number of real-time examples in visual tracking where the performances are signiJicantly improved with the Weighted Bayesian Networks.