University of Washington / Department of Computer Science and Engineering / GRAIL / Projects

Digital Humans



The goal of this project is to create the appearance of realistic human characters. This task can be broken down into many subproblems: animation, body shape, surface appearance, etc. So far, most of our work has focused on the body shape subproblem; specifically, how to use data from 3D scanners to create models of human body shape.


Project members


Body Deformations

As the body moves, its shape changes in complex ways, due to the contraction and relaxation of muscles, bones visible beneath the skin, and various other anatomical features. Often this variation is modelled either by artists using various varieties of deformers, or by anatomical modeling and simulation of the bodily structures involved. Instead, we take an example-based approach, where we build a model of these shape changes using 3D laser range scans of a body in various poses.

We scanned an arm, shoulder, and torso in many different poses, a few of which are shown on the left. We then combined these scans into a common parameterization, so that the shape of the body in any pose can be estimated using scattered data interpolation techniques.

Shown below are a few videos of our model, driven by motion capture data:

"calibration" motion (12 MB)

"track" motion (2.5 MB)

"boxing" motion (9.5 MB)

The details of our parameterization and interpolation techniques are described in our SIGGRAPH 2002 paper.


Body Variation

In addition, we also wish to model how body shape varies between individuals. To this end, we used 250 laser range scans of volunteers in approximately the same pose. We developed an algorithm to fit a common template to each range scan in order to create a common parameterization of the body surface. This allows us to morph between body shapes as shown below:

More significantly, our common parameterization also enables us to analyze the variation in body shape, using techniques such as principle component analysis. One possible application is to use the learned distribution of shapes to synthesize new, random individuals. In the figure above, the individuals outlined in red were randomly generated.

Another application is to relate known characteristics of the scanned individual to body shape. For example, we can learn a simple model of how body shape varies with respect to height and weight, and use this to edit an individual, such as the man outlined in red below:

The details of our correspondence algorithm and applications are in our SIGGRAPH 2003 paper.





The arm dataset from our SIGGRAPH 2002 paper on pose pose variation is available online.



This research is supported by: