Multiple description video coding with un-quantized prediction loop
Ran Nathan, Ram Zamir · 2002
Multiple description (MD) coding allows one to recover compressed video from packet loss, e.g., due to transmission over the Internet. The goal is that each packet will provide a self contained coarse description of a picture block, while the combination of all packets should provide the basis for a finer reconstruction of the block. However, direct application of MD quantization after motion compensation is problematic, because in each description the prediction is based on a different quantized version of the reference frame. Previous work on MD video coding suggested to solve this problem by incorporating an additional coarse quantizer into the prediction loop. This solution may reduce the prediction gain when the correlation between frames is high and the quantization resolution is low. We propose to keep the prediction loop un-quantized, while tracking the offset between the predictions in the joint decoder. We further improve the performance by re-designing the index assignment of the MD quantizer to be robust with respect to variable offsets.