State-of-the-art methods reconstruct three-dimensional (3D) face shapes from a single image by fitting 3D face models to input images or by directly learning mapping functions between two-dimensional (2D) images and 3D faces. However, they are often difficult to use in real-world applications due to expensive online optimization or to the requirement of frontal face images. This paper approaches the 3D face reconstruction problem as a regression problem rather than a model fitting problem. Given an input face image along with some pre-defined facial landmarks on it, a series of shape adjustments to the initial 3D face shape are computed through cascaded regressors based on the deviations between the input landmarks and the landmarks obtained from the reconstructed 3D faces. The cascaded regressors are offline learned from a set of 3D faces and their corresponding 2D face images in various views. By treating the landmarks that are invisible in large view angles as missing data, the proposed method can handle arbitrary view face images in a unified way with the same regressors. Experiments on the BFM and Bosphorus databases demonstrate that the proposed method can reconstruct 3D faces from arbitrary view images more efficiently and more accurately than existing methods.
View on arXiv