Motion-compensated prediction is accomplished by searching the previous picture for a block similar to the block being encoded. The enhanced reference picture selection mode allows the encoder to search more than one picture to find the best match and then use the best suited picture to perform motion-compensated prediction. Reference picture selection can be accomplished on a macroblock level.
The selection of the pictures to be used for motion compensation can be performed in one of two ways. A sliding window of M pictures can be used and the last M decoded, with reconstructed pictures stored in a multipicture buffer. A more complex adaptive memory (not specified by the standard) can also be used in place of the simple sliding window. This mode significantly enhances the prediction, resulting in a reduction in the rate for equivalent quality.
However, it also increases the computational and memory requirements on the encoder. This memory burden can be mitigated to some extent by assigning an unused label to pictures or portions of pictures. These pictures, or portions of pictures, then do not need to be stored in the buffer. This unused label can also be used as part of the adaptive memory control to manage the pictures that are stored in the buffer.