A motion vector prediction method for multi-view video coding

This paper proposes a motion vector (MV) prediction algorithm for increasing the gain of multiview video coding (MVC). In the conventional MVC algorithms, the main feature that brings coding gain is the interview MV estimation and compensation in anchor frames. Although all the other frames (non-anchor frames) also consider both temporal and interview MVs, the interview MVs are usually outnumbered by the temporal MVs. Note that the H.264 schemes encode the MVs in DPCM style and thus the accuracy of MV prediction from neighboring ones is important. When an interview MV is being considered, the conventional method predicts it by neighboring MVs regardless of their being temporal or interview directions, which results in inaccurate prediction and thus rare use of interview MVs. Hence we propose a new MV prediction and encoding scheme especially for the interview MVs. The proposed method first estimates the interview MVs that are not sent to the decoder, from the temporal reference frames. The interview MVs are well estimated in most cases because the depth information does not change rapidly. By using the estimated interview MVs instead of temporal MVs, current interview MV being encoded is more accurately predicted and thus the chance of their use is increased. This results in the increase of coding gain about 0.2dB on average and 0.33dB at maximum.