Residues:

Our basic idea is to replace each depth map as the deviation from the common proxy model, thus needing fewer bits to encode the information. We first compute the residue depth maps by projecting the fitted proxy model to each view for each frame. The residue map for each view is the pixel-by-pixel difference between its input depth map and the proxy depth map. This is the prediction error when the proxy model is used as a common prediction for all views. After this process, each of the $mn$ depth maps is replaced by a residue map losslessly. The residues have a smaller range of values and are less bulky to represent. They are also more correlated spatially and temporally within a view, being the differences with a good approximation of the underlying structure.