JVET-N0436 CE3-3.4.1, CE3-3.4.2, CE3-3.4.3: Results for unified MPM list construction process [F. Bossen, K. Misra (Sharp Labs of America)]
This contribution provides results for the following CE3 experiments: CE3-3.4.1, CE3-3.4.2, and CE3-3.4.3. The experiments evaluate performance of a modified luma intra Most Probable Modes (MPM) list comprising solely angular modes, as proposed in JVET-M0528. A flag is used to indicate if the intra mode is angular or non-angular. For the non-angular modes, a further flag is received to indicate Planar or DC mode. It is asserted that the proposed list construction is a simplification over the current VVC design that features four different MPM lists: a list for zero reference line, a list for non-zero reference line, a list for intra sub-partitions and a list for combined intra-inter prediction.
Test CE3-3.4.1 evaluates the proposed list construction with 6 MPM and shows a coding benefit of 0.11% (all-intra) with no impact on encoder runtime. Test CE3-3.4.2 makes use of the same list construction as in CE3-3.4.1 but also tests N lowest bit-cost modes at the encoder. This results in a larger coding gain of 0.29% with a 3% encoder runtime increase. It is also reported that for class A1 sequences in all-intra configuration the proposed test provides a coding gain of 0.52% with no impact on encoder runtime. In CE3-3.4.3, the MPM list construction process is truncated to a size of 4. This results in a slightly lower coding gain of 0.05%. As requested in the CE3-3 description, all three tests were carried out using a further encoder modification (i.e. ALL_MPM) that evaluates all the candidates in the MPM list (for reference line 0) during the R-D mode decision. The performance of the ALL_MPM tests was measured versus anchors that contain the same encoder change. Tests CE3-3.3.4.1, CE3-3.4.2, CE3-4.3 were also evaluated with that a modification that reduces the number of contexts used for coding angular mode flag. The context reduction removes checking of left and above syntax elements and is enabled by setting the macro REDUCE_CTXT to 1 in the software released as part of CE3 testing. Listed below is the summary of test results:
AI | RA | |||||||||
Y | U | V | EncT | DecT | Y | U | V | EncT | DecT | |
3.4.1 | −0.11% | −0.08% | −0.09% | 99% | 101% | −0.02% | −0.02% | −0.13% | 100% | 100% |
3.4.2 | −0.29% | −0.2% | −0.17% | 103% | 100% | −0.08% | −0.01% | −0.07% | 100% | 100% |
3.4.3 | −0.05% | −0.05% | −0.02% | 97% | 100% | 0.00% | −0.01% | −0.04% | 99% | 99% |
3.4.1+ALL_MPM | −0.13% | −0.15% | −0.08% | 99% | 101% | −0.04% | 0.10% | −0.04% | 100% | 100% |
3.4.2+ALL_MPM | −0.13% | −0.14% | −0.08% | 100% | 101% | −0.05% | 0.07% | −0.04% | 99% | 99% |
3.4.3+ALL_MPM | −0.07% | −0.09% | −0.05% | 99% | 100% | −0.02% | 0.08% | 0.02% | 99% | 98% |
3.4.1+REDUCE_CTXT | −0.11% | −0.07% | −0.08% | 98% | 100% | −0.03% | −0.05% | −0.07% | 99% | 100% |
3.4.2+REDUCE_CTXT | −0.29% | −0.18% | −0.18% | 103% | 101% | −0.09% | 0.01% | −0.18% | 99% | 100% |
3.4.3+REDUCE_CTXT | −0.05% | −0.04% | −0.06% | 96% | 100% | 0.00% | 0.02% | −0.10% | 99% | 100% |
A complexity analysis is also provided, including efficient software implementation details.