Related Work

Most image-based rendering algorithms render new views of real-world objects from example images [13]. The most prominent of these algorithms generate new views of static 3D environments, e.g. [15]. These static geometry techniques do not model changes in shape or texture.

Morphing [2,12], linear interpolation, and radial basis functions (RBFs) each model non-rigid deformations, by producing linear combinations of shape and texture from example images. These methods alone are inappropriate to model poses of articulated figures, since they do not capture important constraints such as constant-length limbs. Darrell [8] has demonstrated a clustering algorithm for selecting appropriate example images for pure non-rigid interpolation. Although our work requires more modeling and image analysis, it requires far fewer source images and can potentially extrapolate over a much wider parameter space. Our work could also be used as the rendering module together a clustering method.

Rigid and non-rigid methods have been combined to model human faces [1,18,5]. In these systems, the overall rotation and translation of the face is modelled rigidly, and the local deformations of shape is modeled non-rigidly. Our work is similar. We model each segment of the body with a separate rigid transformation. View Morphing [20] also separates a single rigid motion from changes in shape. This philosophy of our work is also similar to Debevec et al. [9], where an assisted rigid analysis is combined with view-dependent texturing to generate new views of architecture.

Teichmann [23] creates articulated skeletons from a single 3D dataset.