To estimate the configuration of a figure in a captured image, a silhouette image of the figure is scanned to create a signed distance image. This image identifies the distance of each pixel in the image to the closest edge of the silhouette, and indicates whether the pixel is inside or outside of the silhouette. Multiple distance images of this type are employed to generate an eigen-points model, which provides an affine mapping from the signed distance images to the limb parameters of an authored skeleton. When a new input image is received, it is first processed to create the signed-distance image, and this image is applied to the eigen-points model to estimate limb parameters, such as the locations of various joints in the figure. From this information, each foreground pixel in the captured image can be assigned to one of the limbs.