H.264 supports a so-called multi-reference-frame prediction technique.We adjust its three relevant parameters (the number of previous reference frames

the number of skipped frames

and the number of B-frames)

and use their different combinations to encode various test video sequences respectively.Experimental results show that multi-reference-frame prediction does not improve coding performance but involve heavy computing cost in most cases.We propose that the advantage of this technique depends on a sort of video content self-similarity

which seldom occurs in practice.We then suggest not to use the multi-reference-frame prediction technique in most applications.