A statistical pose model is developed for efficient 3D visual tracking of beating heart. The Region of Interest (ROI) on heart surfaces is first pre-tracked with a conventional high-order thin plate spline model. The 3D pose data of the ROI extracted from the pre-tracked results are then used to train a low-order 3D pose model based on the principal component of these pose data. The low-order model is accurate, robust, and efficient for tracking subsequent heart motion as heart beats are quasi-periodic with stable statistics and the redundant degree of freedom for fitting the poses of heart surface is significantly decreased by the principalcomponent-based dimensionality reduction. The proposed 3D pose modeling is validated on the stereo-endoscopic videos recorded by the da Vinci® surgical system.