 |
 |
|
Face shape recovery
Estimating 3D shape from one image is a challenging problem that has attracted sustained research for over four decades. Recovering facial shape is particularly alluring because of the obvious application of 3D face shape information in the domain of face recognition. There is also interesting psychology results that suggest shape-from-shading plays some part in human face processing. Inspired by the success of statistical methods in the areas of facial appearance modelling, the main thrust of our research has been to investigate how statistical models of face shape can be incorporated into a shape-from-shading framework. This work may be divided into a number of avenues. The most conceptually simple approach is to try and learn directly the relationship between face images and face surfaces. To do so, we have developed a couple model framework in which we construct separate models for images and surfaces and then build a third model that learns correlations between the two. This approach is extremely efficient and is capable of recovering face surfaces that are qualitatively good. However, this approach does not make any attempt to model the image formation process and is hence unable to model variations in illumination and reflectance. As an alternative, we have shown how statistical models can be built in the domain of surface normals. This requires overcoming the problem of modelling data which lies on a spherical manifold. We have developed two approaches to this problem. One uses a projection from the cartography literature while the other is based on techniques from differential geometry. We are able to incorporate this model into an iterative shape-from-shading framework which enforces strict irradiance constraints provided by Lambert’s law. The approach is able to recover fine surface detail (such as wrinkles) and has been extended to explicitly account for cast shadows and albedo variations. Finally, we have considered the problem of using more complex reflectance models which more accurately capture the reflectance properties of skin. We have developed a novel framework for shape-from-shading which is able to minimise brightness error functions of arbitrary complexity. When combined with a statistical constraint on the surface height function, these two constraints allow accurate facial shape and reflectance information to be extracted from single grayscale or colour images. Contact: Will Smith  
|