Back to Search Document details
13th Meeting: Marrakech, January 2019 2019-01-19 00:53
CE4: Modification on History-based Motion Vector Prediction
Abstract
This contribution presents 3 changes on History-based Motion Vector Prediction (HMVP) mode [1] and pairwise candidates not using HMVP predictors which reduce the complexity in both encoder side and decoder side. For the “VTM3.0 as reference” tests in RA case with the size of merge list equal to 6, the presented method with HMVP buffer size equal to 10, the sub test a) “first 3 HMVPs pruned by left and above spatial candidates” achieved 0.03% loss in terms of average luma BD-rate; the sub test b) “first 2 HMVPs pruned by left and above spatial candidates” achieved 0.03% loss in terms of average luma BD-rate with 59% time saving for HMVP table update and adding HMVP candidates in merge list in the decoder side.
JVET-M0126 CE4: Modification on History-based Motion Vector Prediction [Y. Han, W.-J. Chien, C.-C. Chen, H. Huang, C.-H. Hung, Y. Zhang, M. Karczewicz (Qualcomm)]

HMVP table without pruning

In this conurbation, redundancy check is removed from HMVP table update.

Subsampling HMVP table when adding HMVP candidates in merger list

In this contribution, the size of HMVP table is increased to 10, and one out of every 3 is picked to add in the merge list. At most 4 HMVP candidates can be used.

In order to reduce the number of pruning, partial pruning is used when adding a new HMVP candidate in merge list. In the sub test CE4.1.2.a, only the first 3 HMVP candidates are pruned by the left and above spatial candidates. In the sub test CE4.1.2.b, only the first 2 HMVP candidates are pruned by the left and above spatial candidates.

HMVP in AMVP mode

The size of HMVP table is increased to 10, and one out of every 3 HMVP candidates is picked to add in the AMVP list.

Remove dependency between HMVP and pairwise candidates

In VTM3.0, pairwise candidates are derived from the candidates in the merge list which brings dependency between pairwise predictors and HMVP predictors. In this contribution, pairwise candidates are generated not using HMVP predictors.

Decisions
merged
In VTM3.0, pairwise candidates are derived from the candidates in the merge list which brings dependency between pairwise predictors and HMVP predictors. In this contribution, pairwise candidates are generated not using HMVP predictors.
Citation