| Summary: | 3D Face reconstruction is the process of estimating the full 3D geometry of a human's face from one or more images. Applications of 3D face reconstruction span many areas, from personalisation of video games and trying on accessories online, to measuring emotional arousal for psychological studies and in medicine, such as simulating the result of reconstructive surgery. Approaches to 3D face reconstruction generally depend on a 3D Morphable Model (3DMM) - a parametric model, where the shape, pose and expression can be adjusted using a small number of parameters. While methods based on such techniques can work well on frontal images, they often begin to fail on cases of large pose, difficult expression, occlusion, and bad lighting. Additionally, encoding detail in so few parameters is not possible.
In this thesis, we propose a novel approach to the problem of 3D face reconstruction: Volumetric Regression Networks. Our non-parametric approach constrains the problem to the spatial domain using an end-to-end network which directly regresses the 3D geometry using a volumetric representation. This avoids the need for 3DMM generation, which involves finding correspondence between all vertices of all training samples, but also the fitting stage, which requires solving a difficult optimisation problem. We demonstrate that doing so can not only provide state-of-the-art results, but also be adapted to other deformable objects, such as the full human body.
|
|---|