Revisiting Bayesian Blind Deconvolution