Dual generative models for human motion estimation from an uncalibrated monocular camera