Offset temporal motion vector predictor (TMVP)

Abstract

An improved method for temporal motion vector prediction for inter block HEVC is provided that relies on a block translational model. The method adds an offset to a temporal motion vector predictor (TMVP) to improve prediction accuracy. The method first designates a current prediction block as an area for motion compensation where all the pixels inside the prediction block perform identical translation temporally using motion vectors MVs. A coordinate offset is then derived for a current prediction block from the MVs of its spatially neighboring blocks. The offset TMVP is then defined for the current prediction block as the MV of an offset block which is in the geometrical location of the current prediction block coordinate plus the coordinate offset in a specified temporal reference picture. The offset TMVP is then used to code MVs.

What is claimed: 
     
       1. A method of temporal motion vector prediction for inter block coding in High Efficiency Video Coding (HEVC) that relies on a block based translational model, the method comprising:
 designating a current prediction block as an area for motion compensation using HEVC where all the pixels inside the current prediction block perform identical translation temporally using either one or more motion vectors MVs;   deriving a coordinate offset for the current prediction block from the MVs of its spatially neighboring blocks;   defining an offset of a temporal motion vector predictor (TMVP) for the current prediction block as the MV of an offset block which is in the geometrical location of the current prediction block coordinate plus the coordinate offset in a specified temporal reference picture; and   using the offset TMVP to code MVs,   wherein the motion vectors of neighboring prediction blocks to the current prediction block are used to calculate the offset for the TMVP,   wherein the neighboring prediction blocks located in a first three positions in a merge candidate list for the current prediction block are used in calculating the offset for the TMVP, wherein the three neighboring prediction blocks comprise a left (L), an above (A), and an above-left (AL),   wherein with the three neighboring prediction blocks, the offset for the TMVP for the current prediction block is derived as median of motion vectors of these neighbors, as follows:
   dx=median (Lx, ALx, Ax) 
   dy=median (Ly, ALy, Ay) 
   wherein Lx, ALx, Ax are the x component of motion vectors of Left neighbor, Above-left neighbor, and Above neighbor, respectively, and   wherein Ly, ALy, Ay are the y component of Left neighbor, Above-left neighbor, and Above neighbor, respectively.   
     
     
       2. The method of  claim 1 , wherein in the step of defining the offset TMVP, the offset TMVP is further defined by the steps comprising:
 assuming that the current prediction block is at the position of coordinate (x, y) in the current picture; and   adding a coordinate offset of (dx, dy) to the coordinate (x,y) to provide the offset TMVP (x′,y′) as
   (x′, y′) =(x, y) +(dx, dy) =(x+dx, y+dy).
 
   
     
     
       3. The method of  claim 1 , wherein syntax elements are required in coding a bitstream to indicate values for the offset TMVP. 
     
     
       4. The method of  claim 1 , wherein one offset is shared for multiple prediction blocks or coded with coarser granularity than the final fractional motion vector accuracy. 
     
     
       5. The method of  claim 1 , wherein motion vectors of neighboring prediction blocks are normalized to compensate for the difference in temporal distances between references used among the neighboring prediction blocks. 
     
     
       6. The method of  claim 1 , wherein the step of using MV prediction with the TMVP to code MVs, comprises:
 providing a TMVP offset mode that can be turned on or off, wherein in the on mode the offset TMVP with the added offset is used to code MVs, and in the off mode a TMVP without the added offset is used to code MVs, wherein an explicit signal is provided for turn on and turn off of the offset TMVP mode using a flag provided at a CU, slice or sequence level.   
     
     
       7. A method of temporal motion vector prediction for inter block coding of video within a bitstream that relies on a block based translational model, the method comprising:
 (a) designating a current prediction block as an area of said video for motion compensation where all the samples defined by the current prediction block perform identical translation temporally using at least one motion vector;   (b) deriving a coordinate offset for the current prediction block from at least one motion vector of a set of spatially neighboring blocks relative to the current prediction block;   (c) wherein a selection of said set of spatially neighboring blocks is based upon a merge candidate list for said current prediction block;   (d) wherein said merge candidate list includes a left neighboring block (L), and an above neighboring block (A), where said merge candidate list includes an ordered list of candidate neighboring blocks, where said left neighboring block (L) is directly left of and aligned with the current prediction block, where said above neighboring block (A) is directly above of and aligned with the current prediction block;   (e) wherein said left neighboring block (L), and said above neighboring block (A), are included within two of a first position, a second position, and a third position of the ordered list of candidate neighboring bocks of said merge candidate list, where said left neighboring block (L) is included within one of said first position, the second position, and third position, where said above neighboring block (A) is included within a different one of said first position, the second position, and third position than said left neighboring block (L);   (f) defining an offset temporal motion vector predictor (TMVP) for the current prediction block as the coordinate offset based upon the at least one motion vector of an offset block selected from the merge candidate list further based upon another coordinate offset in a specified temporal reference picture that is in a temporally different picture than said current prediction block; and   (g) decoding said current prediction block based upon said offset TMVP.