1
Files
facial-animation-using-inve…/chap/chapter4.tex
2023-11-09 20:05:29 +01:00

89 lines
4.4 KiB
TeX

\chapter{Performance Driven Facial Animation}
\label{chap:performancedrivenfacialanimation}
Facial animation is rarely a fully manual process,
either for efficiency reasons as manual animation is a work-intensive process,
or to capture more true-to-life facial expressions,
or in case of real-time applications,
where manual animation is not applicable at all.
Classical face tracking used in film or video games is typically marker-based optical tracking,
but with the emergence of head-mounted technology like virtual- or augmented-reality headsets,
this method is no longer fitting,
because a large part of the face is obstructed when wearing head-mounted displays.
\section{Stereo/Optical Facial Tracking}
\label{sec:facialtracking}
Marker-based face tracking (or motion capture in general) works similar to stereo 3D-scanning:
At least two cameras point at the face to be tracked,
which is augmented with \textquote{markers} (easily identifiable points),
so facial features can be reliably tracked by all cameras.
Because cameras can \textquote{see} the same marker from multiple perspectives (the camera positions have to be known),
the distance of a marker from both cameras can be calculated by solving the intersection of the straight lines defined by each camera position and the marker position.
In the schematic in \autoref{fig:2dstereotracking}, the cameras are marked blue, while the marker on the face is marked red.
\begin{figure}[h]
\centering
\begin{subfigure}[b]{0.75\textwidth}
\begin{tikzpicture}[scale=3]
\clip (-2, -0.2) rectangle (3, 1.3);
% Coordinates
\coordinate (CA) at (-1, 0);
\coordinate (CB) at (1, 0);
\coordinate (M) at (0, 1);
% Axis
\draw[->] (-1.5, 0) -- (1.5, 0); % x
\draw[->] (0, 0) -- (0, 1.3); % y
% Cameras
\node[isosceles triangle, isosceles triangle apex angle=30, draw, fill=blue, minimum size=0.5, rotate=225] (TA) at (CA){};
\node[isosceles triangle, isosceles triangle apex angle=30, draw, fill=blue, minimum size=0.5, rotate=315] (TB) at (CB){};
% Camera beams
\draw[->, blue] (CA) -- (M) -- +(0.2, 0.2);
\draw[->, blue] (CB) -- (M) -- +(-0.2, 0.2);
% Marker
\draw[fill=red] (M) circle (0.025);
% Face
\draw[green] (M) +(0, 0.05) circle (0.2);
\draw[green] (M) +(0.075, 0.1) circle (0.04);
\draw[green] (M) +(-0.075, 0.1) circle (0.04);
\draw[green] (M) +(-0.075, -0.075) arc[start angle=225, end angle=315, radius=0.1]{};
\end{tikzpicture}
\end{subfigure}
\caption{2D Stereo Tracking.}
\label{fig:2dstereotracking}
\end{figure}
When basic information about the camera is known (e.g. focal length and sensor size),
the beam directions can be inferred from the marker's pixel position in the image plane.
This allows defining straight line equations for both beams,
equating those and solving the resulting linear system leads to the distance from the camera.
Because the camera positions are known, the marker's position is also known.
By adding more cameras to the tracking setup and thus increasing the size of the linear system to be solved,
errors can be reduced and the tracking area increased (surround cameras also enable tracking rotations).
Proprietary tracking systems like OptiTrack~\autocite{optitrack} use techniques like infrared cameras with infrared emitters and reflective markers
to lessen the influence of regular daylight and simplify the marker detection.
Finally, when the marker positions have been calculated,
the inverse kinematics problem for either blend shapes or a skeleton can be solved,
since each marker corresponds to a vertex of the facial model.
\section{Tracking Solutions in Consumer VR Products}
\label{sec:consumertrackingsolutions}
Because head-mounted-displays hinder the application of facial markers,
facial tracking data for a portion of a face is typically extracted from a single camera feed instead of multiple.
Cameras might be mounted inside the HMD, facing the eyes and nose, and below the HMD, facing the mouth and chin.
Since a delta blend shape model linearly combines locally isolated deformations,
processing different facial regions individually does not impose a problem.
Unlike the stereo tracking, this approach does not use the inverse kinematics method described in \autoref{sec:blendshapeinversekinematics}.
Instead, each camera feed is fed through a specialized computer vision model,
which directly calculates the blend shape weights for a particular facial region and blend shape model.