Fast Transcoding of Compressed Bitstreams by Reusing Incoming Motion Vectors

Transcoding will be an important activity in the digital studio. People will transcode MPEG-2 bitstreams to resize images, support progressive or interlaced receivers, optimize for editability or storage/transmission efficiency, etc. Transcoding involves decoding a bitstream, possibly modifying the decoded pictures, and re-encoding. Transcoding is expensive because encoding is expensive, and the major cost of encoding is motion estimation (ME). A program, Motion Vector Reuse by Extrapolation (MVREX), has been developed that reuses incoming motion vectors during transcoding. MVREX assumes that motion is linear, and extrapolates known motion from input macroblocks to output macroblocks (MBs), avoiding ME. MVREX has transcoded both progressive and interlaced video between different group of picture (GOP) structures, and supports all MPEG-2 prediction modes for frame pictures, except dual prime. In the cases examined, MVREX produces video with peak signal-to-noise ratio (PSNR) within 1 dB of that produced by a full search, but with orders of magnitude less computation. MVREX continues to work well after multiple generations of transcoding and appears to be a promising approach to transcoding. After a brief review of MPEG-2, this paper describes the concept of motion extrapolation, derives basic formulas, and presents several results and observations.

Print ISSN

0036-1682

Published

1999-08

Content type

Original Research

DOI

10.5594/J15678