This final chapter is on experiments composed of the complete pose- and view-dependent model. Experiments range from qualitative evaluation on simple one degree of freedom bones to reanimation of legs. Other applications include free-viewpoint rendering and acquisition and modeling of an accordion.
Repeatability
In this first qualitative experiment, a robotic arm was clothed and posed to similar positions to test repeatability of the cloth deformations as a function of pose. For tighter fitting clothing the deformations appear to be kinematic.
Face, single arm, and upper body
The following overview video gives results for the face, single arm, and upper body examples.
Robotic Arm
The input for the robotic sequences comes from a moving robotic arm in front of a single camera. The robotis calibrated beforehand (e.g., WamCal2), and the following sequence from a single viewpoint is reorganized to obtain severl virtual viewpoints that observe a small motion.
An illustration of the converted virtual cameras (30+) that observe a small range of motion
Using the above sequence three models were obtained.
Results for the plaid sequence
Results for the corduroy sequence
Results for the PJ sequence
Legs
In order to obtain and transfer cloth deformation on legs, two sets of input data were acquired. The first set of data is acquired with knees bending and can be used to extrapolate the leg motion to e.g., walking up stairs. This was done for two different types of clothing. Notice the wrinkles on the knee change as the pose of the legs changes.
Khaki stair example
Gray stair example
Walking example
In the walking example, due to capture constraints, a simpler leg motion was used during acquisition and was mapped to the novel walk cycles that were created from a physical simulation.
Accordion
The pose- and view- dependent model can be used to model objects other than humans. In this example, three distinct acquisitions of an accordion in three states were obtained to create a pose- and view-dependent model. The accordion bellows have view-dependent effects that are best recreated when the model also takes into account pose dependency
Free-viewpoint compression
The compact model can also be used for compression applications. In the following example, two data sets from MIT are acquired and compressed through the model. The renderings are obtained in the same framework as the other pose- and view-dependent models.
Interpolation
One main consideration when building a model is the interpolation method used. The following video illustrates some of the issues that can occur due to inappropriate interpolation. For example, local RBF's (e.g, a Gaussian) can have too small of support for interpolation. This is illustrated on the face, and can easily be resolved by using a bigger support, by using a global RBF, or by using a linear interpolation.
Nearest interpolation can give good renderings for static images, but will create poor animations. This is illustrated on the walking training data.