Heuristic Background depth maps for each frame can then be computed
by an inverse projection from the panorama depth map using
the previous homography. By design, these depth maps are temporally
stable and consistent throughout the shot. Then, segmentation
of the players is improved considering the background panorama,
and each segmented player is represented as billboard with depth
derived from its location on the background model. Ambiguities in
segmentation are corrected so as to not cause noticeable artifacts,
giving us a final depth map for each input image. Finally, stereo
views are rendered with disocclusions inpainted from known background
pixels. In the following sections we describe each of these
steps in more detail.