Organisation internationale de normalisation

17.7Motion Compensation and Interpolation Filters

17.7.1.1.1.1.1.1.1JCTVC-D138 Rounding control for chroma interpolation [B. Li (USTC), J. Xu (Microsoft), G. Sullivan (Microsoft), F. Wu (Microsoft), H. Li (USTC)]

The relevance of this document was affected by the planned change of the chroma motion interpolation filter to no longer use bilinear filtering. Further study was encouraged to determine whether the proposal could lead to an improvement in a future version.

A cross-check was provided in JCTVC-D310.

17.7.1.1.1.1.1.1.2JCTVC-D175 Bi-prediction combining template and block motion compensations [C.-L. Lee, C.-C. Chen, Y.-W. Chen, M.-H. Wu, C.-H. Wu, W.-H. Peng, H.-M. Hang (NCTU/ITRI)]

This contribution proposed a bi-prediction scheme that combines motion vectors (MVs) found by template and block matchings based on parametric OBMC. Because the template MV is inferred on the decoder side, this technique only sends motion parameters for one block MV. The concept is further extended to allow the template MV to be replaced with one of the MVs for neighboring prediction units, enabling an implementation described as low-complexity and template-matching-free. Experiments were conducted using the TMuC-0.9_HM software and the common test configurations to evaluate three variants of this scheme. The best of them reportedly achieved an average BD BR saving of 2.2%, with a minimum of 1.4% and a maximum of 4.3%. It was reported that the average decoding time increased by 10% while the encoding time doubled.

Some significant gain appeared to be shown, and further study was encouraged (e.g., in a CE).

17.7.1.1.1.1.1.1.3JCTVC-D225 Directional interpolation filters for luma samples using 8 filter coefficients [A. Fuldseth, G. Bjøntegaard (Cisco)]

This contribution relates to improving DIF using 8 tap filtering rather than 6 tap filtering. Due to the decision not to use DIF in the HM at this time, it was not discussed in detail.

For LC RA, an improvement of 2.4% was reported relative to the anchor.

For LC LD, an improvement of 2.6% was reported relative to the anchor.

17.7.1.1.1.1.1.1.4JCTVC-D246 New interpolation filters [J. Lou, L. Wang (Motorola Mobility)] (missing prior, uploaded 20th, before meeting)

DCT-based interpolation filters and their low complexity optimization were reportedly introduced by Samsung in the current TMuC0.9. The 12tap DCT-IF has been used for the Low-Delay High Efficiency tests and the Random-Access High Efficiency tests. This contribution proposes different DCT-based interpolation filters. Compared to the current interpolation filters in the TMuC, better Rate-Distortion performance could reportedly be achieved with the same tap length, or lower complexity could be achieved with the same Rate-Distortion performance.

The proponent indicated that their proposal may provide some improvement in coding efficiency relative to the scheme planned for adoption at this meeting – e.g., 0.1-0.2% for 8-tap filtering. Some participants remarked that the 6 bit precision of the tap values for the adopted scheme was a beneficial property and that shift-add implementation should be considered.

Further study was encouraged (e.g., in a CE).

17.7.1.1.1.1.1.1.5JCTVC-D266 Investigation on Luma Interpolation Filter [V. Sze, M. Zhou, M. Budagavi (TI)]

This proposal advocated the adoption of an 8 tap DCTIF MC interpolation filter, which had been agreed prior to its review, and therefore no detailed review of the contribution was necessary. The contribution advocated this action based on hardware and software complexity analysis (including memory bandwidth) and the amount of coding efficiency impact.

17.7.1.1.1.1.1.1.6JCTVC-D285 Switching Interpolation Filter (SIF) scheme with low-complexity decision algorithm [T. Yoshino, S. Naito (KDDI)]

In the previous meeting, a Switching Interpolation Filter (SIF) approach was proposed to improve coding efficiency. However, the SIF approach with selecting the best filter based on RDO reportedly requires a significant computational cost. In this contribution, a different approach for determining filter coefficients was described, and the coding performance of the proposed SIF scheme with the modified filter decision was discussed. An 8-tap filter framework was employed to evaluate the performance under HE operation. The experimental results reportedly showed BD BR Y to improve 0.4% on average against TMuC0.9 under LD/LC condition while maintaining the encoding/decoding time to be approximately unaffected.

It was noted that the modification of the LC configuration motion interpolation already planned for adoption at this meeting would provide some benefit.

The proponent suggested that the possibility to select from several candidate filters may provide further improvement, although some modification of this proposal would likely be needed to achieve an improvement relative to the new anchor. Further study may result in a new method to provide additional gain.

17.7.1.1.1.1.1.1.7JCTVC-D321 High precision bi-directional averaging [K. Ugur, J. Lainema, A. Hallapuro (Nokia)]

The current low complexity configuration includes rounding control for bi-prediction to combat the rounding error accumulation happening in averaging two predictions. This proposal argues that the error accumulation problem actually doesn’t exist if the rounding operation for a bi-predictive signal is performed at the last step. After interpolating each one of the predictions, this proposal postpones some or all of the rounding to be after both predictions are added, while keeping all the intermediate calculations in 16-bits and does not use the "rounding control" scheme.

It was remarked that some clipping is also needed, and it was asked whether the clipping should be before or after the addition. The tested usage was to clip before the addition.

The decoding time reportedly increased by approximately 5%, which was described as being a consequence of the way some of the processing was implemented in the HM software rather than an intrinsic consequence of the proposed method. It was reported that the software might actually run faster if this part of the processing were changed.

It was noted that in AVC, weighted prediction is another way to introduce an offset in the decoding process.

The proponent indicated that, if used, the technique should be applied to chroma as well as luma.

Decision: Adopted into the HM.

17.7.1.1.1.1.1.1.8JCTVC-D352 High accuracy averaging for bi-prediction (Verification of JCTVC-D321) [F. Bossen, S. Kanumuri (DOCOMO USA Labs)]

Cross-verification of JCTVC-D352 with some additional information provided. Careful study and analysis of the proposal was reported, as well as checking of results.

17.7.1.1.1.1.1.1.9JCTVC-D397 A New Adaptive Interpolation Filtering Technique [F. Kossentini, N. Mahdi, M. A. Ben Ayed (eBrisk Video Inc)]

In this contribution, an Adaptive Interpolation Filtering (AIF) technique using adaptive filters was proposed for interpolation filtering of the luma samples of video sequences. The proposed technique allows the encoder to switch, for each of the three groups of positions, between the TMuC 0.9 fixed filter (12-tap and 6-tap for High Efficiency and Low Complexity profiles, respectively) and a newly-generated AIF filter. Compared to the TMuC 0.9, the proposed technique yields average BD BR reductions of 0.6%-0.9% for the two Low Delay configurations and the Random Access / Low-Complexity configuration.

12 tap filters were used for the HE case.

6 tap filters were used for the LC case.

The encoder selects, on a frame basis, whether to use the built-in fixed filters or to use tap values that are transmitted by syntax.

The fractional positions are classified into categories, and the encoder can indicate to use the built-in filters for some categories and transmitted filters for the others.

A coding efficiency improvement of 0.6-0.9% was reported for both the HE and LC cases.

The reported results were not complete.

The encoding times in these experiments were increased.

The proponent indicated that it would be possible to constrain the add/shift complexity of the custom filters.

Further study was encouraged (e.g., in a CE).

17.7.1.1.1.1.1.1.10JCTVC-D438 Directional interpolation filter: alternative directions for f, i, k, q [Hongbo Zhu] (late registration Saturday 22nd after start of meeting, uploaded Monday 24th, fifth day of meeting)

This contribution presents alternative directional interpolation filter for sub-pel positions referred to as f, i, k and q.

A reduction of number of operations compared to the DIF in previous HM is reported, and the change in compression performance is said to be negligible.

After the decision made in CE3 (to use a unified interpolation filter, DCTIF-8tap) this proposal appeared not to need further consideration.