JCTVC-M0393 Non-SCE5: Crosscheck of JCTVC-M0144 on combined tests of SCE5.1.5 and SCE5.2.* [T.-D. Chuang, Y.-W. Huang (MediaTek)] [late]
JCTVC-M0112 Non-SCE5: Position derivation for using Base Layer motion in SHVC [C. Gisquet, P. Onno, E. François, G. Laroche (Canon)]
Since HM2.0, memory compression has been adopted so as to reduce the storage requirement of motion data to be used in deriving the temporal MV predictor. Following this, the location of that predictor has been changed to the bottom-right position H of the PU. It is asserted in this contribution that a similar issue occurs when deriving the motion candidate from the base. It is therefore proposed to perform a different rounding on the colocated coordinates within the Reference Layer. The gains of this rounding over SHM 1.0 are reported to be respectively for the TextureRL approach, −0.8%(RA 2x), −0.2%(RA 1.5x), −0.6%(RA SNR) and −0.5%(LD-P 2x), −0.1%(LD-P 1.5x), −0.5%(LD-P SNR). For the reference frame index approach, only ratio 2.0x is reported to offer changes, with gains over SHM1.0 of −0.2% (RA 2x) and −0.1% (LD-P 2x).
The average gain is about 0.5%. Basic idea is that the spatially closest position from the 16x16 MV memory is used. The position taken from the base layer may therefore vary depending on the rounding process (rounding performed in base layer coordinates).
The gain is slightly larger than 5.1.5.2 (which uses a second candidate from bottom right), but the coordinate rounding is simpler than the additional MV scaling necessary in 5.1.5.2 (according to the opinion of cross-checkers).
Combinations with 5.1.5.1 (and other approaches reducing number of comparison in pruning) and with 5.2.2 would be interesting.
Further study in CE
Dostları ilə paylaş: |
