Building a photorealistic and animatable model of the human head by hand is
slow, and existing high-quality approaches rely on multi-camera capture that
remains out of reach for ordinary users. In this thesis we built a complete
pipeline that produces a photorealistic and animatable head avatar from a
single-camera video and drives it live from a webcam. The face is tracked
with VHAP, which estimates FLAME model parameters, and the avatar is
built with GaussianAvatars, which binds Gaussian primitives to the triangles
of the mesh. The preparation and runtime pipelines are separate and meet
only at the FLAME model parameters, so the source of these parameters can
be replaced without retraining the avatar. On the same avatar we therefore
compared a published mapping of MediaPipe blendshape coefficients with our
own hybrid approach based on the SMIRK regressor, to which gaze estimation
is added from the blendshape coefficients. We derived a linear conversion
between the expression bases of the two FLAME versions and transferred
eyelid closure as a direct vertex displacement. We also developed a guided
capture program that steers the user towards even coverage of head poses
and continuously checks sharpness.
The hybrid approach produced noticeably better animation, and the
difference appears exactly at those deformations that the FLAME expression
basis does not capture, most clearly eyelid closure. Driving runs in real
time on a single consumer GPU. We further show that capture sharpness
and even coverage of head poses strongly influence quality, and that the
default densification settings keep increasing the number of primitives during
longer training without improving quality. The usable range of viewing angles
corresponds to the range covered during recording.
|
