Graphics Reference
In-Depth Information
denotes the set of visible cameras, c
V
where
the covariance matrix of the projection of
the Gaussian
α c is the cosine of the foreshortening
angle observed at camera c . The variance of the Gaussian
N
into camera c , and the weighting term
is chosen such that high spatial
frequencies are attenuated. It can either be defined directly on the surface using the known
maximum size of the features or in dependence of the matching window m . The next steps
are based on the assumption that variation in mesoscopic intensity is linked to variation in the
geometry. For human skin, this is mostly the case. Spatially bigger skin features tend to be
smooth and are thus filtered out. The idea is thus to adapt the local high frequency geometry of
the mesh to the mesoscopic field ( X ). The geometry should locally form a concavity whenever
( X ) decreases and a convexity when it increases.
N
1.4 Dynamic 3D Face Reconstruction
The objective now is to create dynamic models that accurately recover the facial shape and
acquire the time-varying behavior of a real person's face. Modeling facial dynamics is essential
for several applications such as avatar animation, facial action analysis, and recognition.
Compared with a static or quasi-static object (or scene), this is more difficult to achieve
because of the required fast processing. Besides, it is the main limitation of the techniques
described in Section 1.3. In particular, laser-based scanners and time-coded structured light
shape capture techniques do not operate effectively on fast-moving scenes because of the
time required for scanning the object when moving or deforming. In this section, we present
appropriate techniques designed for moving/deforming face acquisition and post-processing
pipeline for performance capture or expression transfer.
1.4.1 Multiview Dynamic Reconstruction
Passive facial reconstruction has received particular attention because of its potential appli-
cations in facial animation. Recent research effort has focused on passive multi-view stereo
(PMVS) for animated face capture sans markers, makeup, active technology, and expensive
hardware. A key step toward effective performance capture is to model the structure and
motion of the face, which is a highly deformable surface. Furukawa and Ponce (2009) pro-
posed a motion capture approach from video stream that specifically aims at this challenge.
Assuming that the instantaneous geometry of the face is represented by a polyhedral mesh
with fixed topology, an initial mesh is constructed in the first frame using PMVS software
for MVS (Furukawa and Ponce, 2010) and Poisson surface reconstruction software (Kazhdan
et al., 2006) for meshing. Then its deformation is captured by tracking its vertices v 1 , ...
v n
over time. The goal of the algorithm is to estimate in each frame f the position v f
i of each
vertex v i (From now on, v f i will be used to denote both the vertex and its position.) Each vertex
may or may not be tracked at a given frame, including the first one, allowing the system to
handle occlusion, fast motion, and parts of the surface that are not initially visible. The three
steps of the tracking algorithm refer to local motion parameters estimation, global surface
deformation, and filtering.
First, at each frame, an approximation of a local surface region around each vertex, by its
tangent plane, gives the corresponding local 3D rigid motion with six degrees of freedom.
Search WWH ::




Custom Search