An Efficient Six-Parameter Perspective Motion Model for VVC

• A six-parameter perspective motion model. • Simplified perspective transform. • Modeling keystone motion fields for first- and third-person video. • Advanced perspective motion vector prediction for perspective motion model. • Perspective merge mode for perspective motion model. Tilt and pan camera movements are common in computer games or social media videos. These types of videos contain numerous perspective transforms while today’s video codecs rely on translational and affine motion models for motion compensation. The general perspective motion model with 8 parameters (8PMM) has unreasonably high processing time. In this paper, the eight-parameter perspective transform is simplified into a six-parameter transform to keep the time complexity within an acceptable range while modeling the most relevant transforms. Also, two motion prediction modes, Advanced Perspective Motion Vector Prediction (APMVP) and Perspective Model Merge (PMM), are proposed. The implementation results show an average of 7.0% BD-rate reduction over H.266/VVC Test Model with a maximum of 20% encoding time overhead. The results also show a 71% processing time reduction in comparison to 8PMM while experiencing an average of 5.6% increase in BD-Rate. Much better visual quality is measured through VMAF quality meter.