Compressed Representation:

The data to be sent to the rendering client includes the following. (a) The joint-angle parameters for each frame, (b) the mask bits for each view and each frame, (c) $m$ MPEG streams for the residue values when using MPEG or the sign bit and R orD values for the each frame. We use the following compression schemes for each. Please note that the articulated model and the skinning triangles are available at both ends and need not be transmitted at all.

1. The model parameters (joint-angles) take only few bytes per frame. Only entropy encoding using zip is used to reduce its size.
2. The mask planes are bit-maps. They are compressed using a lossless JBIG format to get maximum compression. Run-length encoding and other schemes were also tried, but JBIG gave the best performance in all cases.
3. When using MPEG, the data is encoded using the ffmpeg library with the given bit-rate.
4. When using bit-plane encoding, the sign bits are compressed using the JBIG scheme. The R and D frames are compressed into a long sequence of bits and then entropy coded using zip.

Figure 6: Results showing input depth, residues, residue differences, sign and reconstructed depth.

Each of the $m$ views are treated independently for this process. The frames of each view are also independent, with some using residue values and others using residue differences. All data is zipped together at the end as a simple entropy-encoding method. This information is sent to the rendering client, which decodes the depths back. The final representation contains one parameter file per frame, one mask, one sign bit, and an R or D plane per frame per view. The data for a whole block (between two R frames for bit-plane coding or between two I frames for MPEG) needs to be together logically and can be treated like a package to be sent to the client.

Figure 7: Results for IndianDance Dataset with block size=25, joint angle noise=5, and $K$=0,5-14, $k$=0-4