International organisation for standardisation

6.7Motion compensation operation and interpolation filters

BoG activity planned (T. Suzuki)

6.7.1Interpolation filters and MV precision

6.7.1.1.1.1.1.1.1JCTVC-F216 An Adaptive Interpolation Filtering Technique [F. Kossentini, N. Mahdi, H. Guermazi, M. A. Ben Ayed, M. Horowitz (eBrisk)]
6.7.1.1.1.1.1.1.2JCTVC-F242 Bilinear chroma interpolation for small block sizes [K. Ugur, J. Lainema (Nokia), K. Kondo, T. Suzuki (Sony)]
6.7.1.1.1.1.1.1.3JCTVC-F468 An Adaptive Interpolation Filtering Technique [F. Kossentini, N. Mahdi, H. Guermazi, M. A. Ben Ayed, M. Horowitz (eBrisk)]
6.7.1.1.1.1.1.1.4JCTVC-F719 Cross-check report for Qualcomm's JCTVC-F585 on Luma/chroma interpolation precision [J. Chen (Samsung)] [late reg. 07-12, upload 07-12]
6.7.1.1.1.1.1.1.5JCTVC-F599 On Chroma interpolation filters [Koohyar Minoo, Jian Lou, Ajay Luthra]
6.7.1.1.1.1.1.1.6JCTVC-F685 Cross-check of JCTVC-F599: Chroma interpolation filters [Jinwen Zan, Dake He] [late reg. 07-07, upload 07-15]
6.7.1.1.1.1.1.1.7JCTVC-F730 Cross-check for Motorola’s Proposal (JCTVC-F601) on Sub-pixel Interpolation [Z. Zhou, S. Liu (MediaTek)] [late reg. 07-13, upload 07-16]
6.7.1.1.1.1.1.1.8JCTVC-F602 Cross-check for combined interpolation filter design (JCTVC-F468) by Samsung [E. Alshina (Samsung)] [late upload 07-12]

6.7.1.1.1.1.1.1.9JCTVC-F601 Joint sub-pixel interpolation for bi-predicted motion compensation [K. Minoo, J. Lou (Motorola Mobility)]

This document proposes an interpolation scheme for temporal prediction of PUs (prediction units), when bi-prediction is used. In such cases the optimal filter per reference list would depend on the joint sub-pixel offsets. For example if four sub-pixel offsets per reference block/list can be indicated for prediction of each PU, then a set of up to 16 pairs of filters can be designed to optimally conduct a bi-prediction motion compensation for that PU. This scheme in intended to be used to increase the coding efficiency of video compression schemes, or alternatively to reduce the complexity of motion-compensation, in terms of memory access bandwidth and/or number of operations performed for sub-pixel interpolation filtering.

Proposes different (likely shorter) filters for biprediction versus uniprediction.

With 8 taps for unipred (as in current design) and 6 for bipred, roughly no loss in coding efficiency was reported.

With further shortening to 4 tap for bipred, some loss was observed.

If the bipred filters are made 8 taps or by using improved 6 tap filters, potentially some coding gain may be observed.

Further study was encouraged.

6.7.1.1.1.1.1.1.10JCTVC-F125 Progressive MV Resolution [J. An, X. Li, X. Guo, S. Lei (MediaTek)]

This contribution proposes a progressive MV resolution (PMVR) method, which allows adaptation among different MV resolutions. In PMVR, MV resolution is not explicitly signaled with specific syntax but progressively adjusted based on the magnitude of MV differences. It is reported that 1.0% luma BD-rate reduction is obtained on average for all four inter configurations (RA-HE, RA-LC, LB-HE, and LB-LC). Moreover, it is also reported that average encoding time is reduced by 5% while average decoding time is increased by 2%. For LP-HE and LP-LC configurations, average luma BD-rate reduction is reportedly 2.8% with similar encoding and decoding times.

Has tuning parameters sent by the encoder, and adjustments that depend on QP of reference pictures.

Requires more interpolation filter kernels, equivalent to 1/8 pel MV support.

A substantial amount of the gain comes from BQSquare, which tends to show benefit from high-precision motion in general. It was remarked that B pictures is another way to get that benefit, as it enables pseudo-1/8 pel motion.

It was asked whether the result had been compared with the previously proposed AMVRES scheme (proposed by Qualcomm).

It was noted that another related proposal had also been submitted.

A participant remarked that other such methods seemed simpler.

The proposal seemed to require a significant amount of pseudo-code to compute MVD.

It was noted that just changing the interpolation filter can help for the P picture cases.

To be further studied in a CE.
6.7.1.1.1.1.1.1.11JCTVC-F567 Adaptive resolution on motion vector difference [W.-J. Chien, P. Chen, X. Wang, M. Karczewicz (Qualcomm)]

This contribution presents a coding method to signal the resolution of the motion vector differences. The motion accuracy of the motion vector difference can be adaptively selected to be 1/4th pel or 1/8th pel and signaled via a motion resolution flag. A joint coding of motion resolution flag and motion vector difference are also proposed. Simulation results reportedly show an average 0.4% BD-rate saving on the high efficiency configurations and an average 0.6% BD-rate saving on the low complexity configurations.

Uses some thresholds that are transmitted.

Gain is about half that reported for JCTVC-F125.

The additional gain may come from the support of a switching point between 1/4 and 1/2 pel in the other scheme.

Roughly conceptually similar to (part of JCTVC-F125) and similar elements to JCTVC-F471.

6.7.1.1.1.1.1.1.12JCTVC-F692 Crosscheck for Qualcomm's Proposal JCTVC-F567 [J. An, X. Guo (MediaTek)] [late reg. 07-09, upload 07-21]
6.7.1.1.1.1.1.1.13JCTVC-F471 Picture Adaptive 1/8-pel Motion Compensation Method [T. Sugio, T. Nishi (Panasonic)]

In this contribution, a picture-adaptive 1/8-pel motion compensation method was proposed. In the proposed method, For P slice, 1/8-pel motion compensation is enabled and all reference pictures are referred using 1/8-pel motion vectors. On the other hand, for B slice, 1/8-pel motion compensation is enabled but all reference pictures are referred using 1/4-pel motion vector in order not to increase overhead to signal motion vector difference. Experimental results reportedly showed 2.5% BR saving for HE and 2.0% BR saving LC on average in LD P scenarios. It also reportedly showed 0.3% BR saving for HE and 0.4% BR saving in the RA scenarios, and 0.2% BR saving for HE and 0.3% BR saving for LC on average in the LD B scenarios relative to the HM3.0.

1/16 pel for chroma when 1/8 pel used for luma.

Difference between LP and LB is reportedly about 6%.

Internal increase of MV precision for predictor.

Biggest difference for BQTerrace, BQSquare, PartyScene.

Closely related to JCTVC-F125 and JCTVC-F567.

6.7.1.1.1.1.1.1.14JCTVC-F682 Cross check for Picture Adaptive 1/8-pel Motion Compensation Method (JCTVC-471) by Samsung [E. Alshina (Samsung)] [late reg. 07-07, upload 07-07]

6.7.1.1.1.1.1.1.15JCTVC-F616 Cross-check of MediaTek’s proposed JCTVC-F125 on Progressive MV resolution [W.-J. Chien (Qualcomm)] [late upload 07-05]