Abstract:
Dynapack exploits space-time coherence to compress the consecutive frames of the 3D animations of triangle meshes of constant connectivity. Instead of compressing each frame independently (space-only compression) or compressing the trajectory of each vertex independently (time-only compression), we predict the position of each vertex v of frame f from three of its neighbors in frame f and from the positions of v and of these neighbors in the previous frame (space-time compression). We introduce here two extrapolating spacetime predictors: the ELP extension of the Lorenzo predictor, developed originally for compressing regularly sampled 4D data sets, and the Replica predictor. ELP may be computed using only additions and subtractions of points and is a perfect predictor for portions of the animation undergoing pure translations. The Replica predictor is slightly more expensive to compute, but is a perfect predictor for arbitrary combinations of translations, rotations, and uniform scaling. For the typical 3D animations that we have compressed, the corrections between the actual and predicted value of the vertex coordinates may be compressed using entropy coding down to an average ranging between 1:37 and 2:91 bits, when the quantization used ranges between 7 and 13 bits. In comparison, space-only compression yields a range of 1:90 to 7:19 bits per coordinate and time-only compressions yields a range of 1:77 to 6:91 bits per coordinate. The implementation of the Dynapack compression and decompression is trivial and extremely fast. It perform a sweep through the animation, only accessing two consecutive frames at a time. Therefore, it is particularly well suited for realtime and out-of-core compression, and for streaming decompression.

Keywords: Animation compression, triangle meshes, out-of-core codec, 3D compression, geometry prediction

You can access this technical report via: PDF