JVET-F0028 EE3: BIO w/o block extension [A. Alshin, E. Alshina (Samsung)]

In EE3 Bi-directional optical flow algorithm was modified in order not to access samples on reference frames other than what regular motion compensation uses. This contribution reports test results of this modification. The reported average performance change is 0.1% in RA and 0.1% in LD-B. This simplification reportedly leads to the 5% encoder and 12% decoder run-time reduction in RA test. For a corner case (44 bi-predicted sub-blocks), the proposed modification reduces the number of multiplications by a factor of two and memory access by factor of three.

In JEM5.0, BIO performs prediction and calculates gradients for the extended block WH (W+4) (H+4). Block extension has been removed. After this modification, the memory bandwidth of BIO is equal to that of regular bi-predictive motion compensation. Computational complexity also has been reduced which results in both encoder and decoder run time reduction.

Questions recommended to be answered during EE tests.

[Q]: Test results with weights not depending on PU size are requested,

[A]: Proposal was tested with all weights equal to 1. Original proposal has no performance drop in RA and 0.1% performance drop for LD. With all weight equal to 1 drop is 0.1% in both R and LD test.

Summary from EE report: With 0.1%performance drop in RA and LD encoder run time can be reduced by 5% and 2%, respectively, decoder run time can be reduced by 12% (RA) and 5% (LD) if BIO doesn’t use block extension. Memory for modified BIO is equivalent to the regular bi-directional MC.

During the discussion, some concern was expressed that the method of padding that is replacing the block extension would disallow sharing gradient computation between adjacent blocks and subblocks with identical motion vectors. However, the worst case memory bandwidth which would occur when all 4x4 blocks have different motion vectors. Also the results clearly indicate an encoder and decoder runtime reduction (confirmed by cross-check).

The description of the method is also consistent and does not cause any mismatch between encoder and decoder, as the padding and subsequent gradient calculation are executed once the block size and the motion vector for the block are known and BIO is invoked.

Decision: Adopt F0028.

JVET-F0028 EE3: BIO w/o block extension [A. Alshin, E. Alshina (Samsung)]