High accuracy motion vectors for video coding with low encoder and decoder complexity

摘要

Two-stage interpolation can be provided for frame prediction samples with quarter-pixel and finer accuracy. All samples of quarter-pixel and finer accuracy can use either half or full/integer-pixels in a bi-linear interpolation to allow for the use of higher accuracy motion vectors, such as one-eighth-pixel accuracy motion vectors. The motion vectors can be restricted in a manner such that they are not allowed to point to every possible sub-pixel sample on a sub-pixel grid, but rather a subset of those sub-pixel samples. In addition, the same full/integer and half-pixel samples that can be used to obtain a quarter-pixel sample can also be used to obtain a one-eighth-pixel sample that the quarter-pixel sample is connected to. Hence, for every quarter-pixel sample, a motion vector could point to two additional positions/locations with one-eighth-pixel accuracy, where the two additional positions/locations can be calculated using the same half or integer-pixel samples which were used to obtain the quarter-pixel sample. Therefore, an arbitrary motion vector accuracy can be achieved without the need to perform additional interpolation stages or increase the complexity of a decoder configured to decode motion vector data.