biprediction

Bidirectional prediction, or biprediction, is a technique in video compression in which the current block is predicted from two reference frames rather than from a single reference. It is most commonly employed in B-frames, which are frames predicted using both past and future references to improve temporal redundancy reduction.

To form a bi-predicted block, the encoder generates two inter predicted blocks: one from a forward reference frame and one from a backward reference frame, using separate motion vectors. These two predictors are combined, typically as a weighted average: P_bi = w1P_fwd + w2P_bwd, with optional offset. The weights may be signaled per block and can reflect lighting differences or motion disparities between the references. The resulting bi-predicted block reduces the residual error that must be encoded.

Bi-prediction enhances compression efficiency by leveraging information from two temporal directions. It is widely used in standards such as H.264/AVC and HEVC for B-frames and B-slices; some codecs allow multiple reference frames and adaptive weighting to better handle occlusions and scene changes. In practice, simple averaging (weights equal to 0.5) is common, but non-uniform weights provide greater flexibility.

While bidirectional prediction improves quality at a given bitrate, it increases computational complexity on both the encoder and decoder due to two motion searches and signaling of weights. It also requires careful handling of interpolation and color/brightness consistency to avoid artifacts when references differ significantly.