BIO performance complexity trade-off

Bi-directional optical flow (so-called BIO) is part of Joint Exploration Model (JEM) which explores potential coding efficiency improvement over state-of-the-art video codec. BIO allows fine motion compensation on a sample level without additional signaling, since refinement is explicitly calculated using just texture information from both reference frames under assumption the validity of optical flow equation. BIO reduces BD-rate in average by more than 2% (up to 5% for some test video), but computational complexity is rather high. Two simplifications for BIO are studied in this paper. First is redesign chain for MC prediction and gradients calculation scheme. Simplified scheme has slightly high latency but reduces amount of multiplications in bi-predicted blocks by factor 2. Another simplification is clustering samples in order to perform motion refinement in BIO not per sample but for group of samples. This allows reduction of division operation in BIO by factor 9.7 in average (up to 256 times in largest blocks). Both modifications enabled together maintain the same performance for BIO in JEM while reduce encoding and decoding run-time significantly.