Automatic facial feature localization plays an important role in many face identification and expression analysis algorithms. It is a challenging problem for real world images because of various face poses and occlusions. This paper proposes a unified framework to robustly locate multi-pose and occluded facial features. Instead of explicitly modeling the statistical point distribution, we use a sparse linear combination to approximate the observed shape, and hence alleviate the multi-pose problem. In addition, we use sparsity constraint to handle the outliers that can be caused by occlusions. We also model the initial misalignment and use convex optimization techniques to solve them simultaneously and efficiently. This proposed method has been extensively evaluated on both synthetic and real data, and the experimental results are promising.