Avatars

We looked at two challenging aspects of creation of avatars of real people. The first aspect is creation of hair styles that will match the look of the users. To do so, it is enough to get images by a camera of the person from different direction. Our system generated a physical (and animate-able) hair structure that is similar to the pictures.  (This was the first work that showed the ability to use multi-view reconstruction, when the hair fibers themselves are too small to be seen in the images, let alone be matched between the views. This work was presented at Siggraph 2005 in the paper “Modeling Hair from Multi View” by Wei et. al. (SIGGRAPH talk)

Figure 1: Automatically generated hair for an avatar from images.

This work was extended to use input from multiple video cameras and recovery of both the hair model as well as motion (“Video-vased modeling of dynamic hair” Yamaguchi et. al.,  PSIVT 2009: Advances in Image and Video Technology pp 585-596).

Another challenge is texturing avatars from original images of the real people. The avatar, constructed from the images had differences in geometry from the original images. Those differences raise from the source of the avatar geometry (Such as a modification of a generic model),  inaccuracies in the geometry reconstruction from stereo, inaccuracies in the position of the cameras, and last but not least, simplification of the avatar geometry to fit real time applications need. Our system, presented at  EuroGraphics 2010 (“Seamles Montage for Texturing Models” Gal et. al.) is able to compensate for all the above elements and generate a consistent texturing of the avatar, even if the geometry of the avatar is quite different than that of the actual person.

Figure 2: A texturing of a model where the smoothed geometry is very different than the actual image (for example, note the missing cavity for the driver in the middle of the car, or the lack of wheels in the model).

        Figure 3: Texturing of a very low level of details model   (used for background avatars) using existing news photography. The model contains no facial geometry, and has large differences than the real geometry, yet the generated texturing is consistent.