Accelerated MPEG Compression of Dynamic Polygonal Scenes

Graphs / Figures

Figure 1 (1172k) and figure 2 (601k) are 24-bit TIFF color images at 1056x720 resolution. Although the actual size used in the compression tests was 352x240, this larger size allows you to more closely examine the details, especially in figure 2.

The graphs, figures 5 through 7, are only a selection of a much larger set of data, presented in full on the CD-ROM and WWW.

The following graphs all use the IPPP frame pattern. The first four graphs show compression on only P-frames. The next four graphs show total compression for the whole scene. The technique described in the paper helps the P frames, and has no effect on I frames.

These graphs explore texture mapping vs. flat shading of the polygons as well as stationary objects (only camera motion) vs. moving objects.

The following graphs all describe moving objects with texture mapping. These files explore using different compression patterns. You'll notice B-frame benefit less from the acceleration technique. Using the default parameters to the MPEG coder, B-frames are heavily quantized, which increase overall scene compression. This also means there's little benefit from a better prediction, since most of the information is thrown out by the quantization.

In all cases, the graphs describe compression vs. search range. Compression refers to uncompressed 24-bit images in and MPEG out. So, a compression ratio of "48" means "half a bit per pixel". A search range of N defines a 2Nx2N range of pixels around the origin. Using "half-pixel" searching, this means an "exhaustive" search of range 10 will consider roughly 1600 possible motion vectors per block.

For more details on what these parameters mean, please refer to the manual which accompanies the Berkeley MPEG encoder.

Dan Wallach, CS Department, Princeton University