3D Facial landmarking plays an important role on 3D face recognition and face expression recognition. However, the most of methods underperform when faces have occluded region such as hair, glasses or finger. To solve this problem, a coarseto-fine method is proposed, containing several denoising auto-encoder networks (denoted as DANs). DANs not only can recover the lost information but improve the accuracy of landmarking. Tests based on Bosphorus dataset show a 100% of good landmarking under 6mm precision of mean error, which demonstrates that our algorithm achieves the state-of-theart performance.