Region based motion vector prediction using data hiding and decoder side reasoning

Motion information occupies a considerable portion of the bit stream generated by the state-of-the-art video codecs. As a sequence, reducing bits on motion information can improve the overall performance of video compression. Various competition based motion vector coding (MVC) schemes were proposed for this goal. In this paper we propose a region based motion vector prediction (RMVP) algorithm to save signaling bits in conventional MVC schemes. Specifically, the motion vector predictor selection information is embedded into the parity of the motion vector differences which would be transmitted into the decoder. To obtain the modulated motion vector differences, the encoder needs to divide the reference frame into several regions and then perform constrained motion estimation in some of the regions. To confirm the advantages of RMVP over different MVC schemes, we implemented it on both KTA Software and TMuC test model. Experimental results show that RMVP outperforms existing MVC schemes.