SignatureBA 017
AuthorMichael Andersch
TitleLossy Parametric Motion Model Compression
TutorDipl.-Ing. Michael Tok, Dipl.-Ing. Alexander Glantz
ProfessorDr.-Ing. Thomas Sikora
AbstractIn contemporary video codecs, motion compensated prediction is the main technique used for reducing inter-frame redundancy. Most codecs describe motion with simple translational models using motion vectors. More sophisticated motion such as rotation or zoom require motion models with more parameters, so called parametric motion models, to be adequately modelled. These additional parameters, however, must be transmitted along with the video sequence, causing significant bit rate overhead. Therefore, it is necessary to compress the additional motion parameters.
In this work, compression schemes for highly precise 8-parameter parametric motion models, so called perspective motion models, are investigated. For this purpose, a multitude of compression strategies is developed and tested using a specialized testing framework. Among these strategies, the most promising one is selected and integrated into a state-of-the-art coding environment based on the in-development video codec HEVC. The performance is evaluated for two different parametric motion model applications running on top of HEVC.
During compression strategy development, the previously best approach to compression of parametric motion models, which operated on 6-parameter affine motion models, is significantly outperformed. Integration of the newly developed compression scheme into the HEVC applications shows gains in bit rate up to 2.314%.
Key wordsparametric motion model, compression, HEVC, parameter, video codec, coding,