International organisation for standardisation organisation internationale de normalisation

18.3Motion compensation and interpolation filters

JCTVC-C126 Chroma interpolation filtering using high precision filter [D. Y. Kim, J. P. Kim, Y. L. Lee (Sejong University), J. Lim, J. Song (SKT)]

In the TMuC, chroma interpolation is performed with the bilinear interpolation to obtain the 1/8-pixel vectors in the same way as in AVC. In this contribution, a chroma interpolation method using the 6-tap AVC luma MC interpolation filter with high precision was proposed. Experimental results reportedly show that the average BD-rate improvements on chroma U and V components are 12.2% and 13.4%, respectively, in the random access, low complexity case, and the average BD-rate improvements on U and V components were reported as 4.1% and 5.0%, respectively, in the random access, high efficiency case. The average BD-rate improvements on U and V components was reported as 12.0% and 14.9%, respectively, in the low delay, low complexity case, and the average BD-rate improvements on U and V components was reported as 2.0% and 2.4%, respectively, in the low delay, high efficiency case. The average BD-rate improvements on Luma were reported as 1.3% and 0.4%, respectively, in low complexity and high efficiency of the low delay case, and the average BD-rate improvements on Luma were reported as 0.2% and 0.4%, respectively, in low complexity and high efficiency of the random access case. Remarks recorded during the review included:

• 
  The results are similar to what has been reported before in TE12 and other TEs.
  
• 
  Establishing a CE on luma and chroma subpel interpolation was recommended
  
• 
  It is still unclear what the real benefit in terms of overall BR gain is.
  
• 
  Complexity is certainly increased compared to bilinear interpolation.
  
• 
  Subjective quality may also be an issue in this context.
  

JCTVC-C137 Decoder-side block boundary decision (DBBD) with OBMC [S. Fukushima (JVC)]

This proposal focused on the tool "Decoder-side Block Boundary Decision (DBBD)" which was proposed in JCTVC-A108 as one of tools for HEVC CfP response. DBBD aims to perform motion compensation with various block patterns without additional coding bits by deciding the MC boundary on the decoder-side. In this proposal, DBBD with OBMC was proposed for improving the quality of MC block in addition to the previous proposal. The simulation results reportedly show that the proposed technique provides 3.1% BD bit rate gain for random access, and 2.2% for low delay as against JM16.2 anchor under HEVC common test conditions. Remarks included the following:

• 
  Only used in 16x8 and 8x16 modes
  
• 
  Encoding time increased to roughly 400-500% (Currently ME is performed for every possible boundary position). Decoder time increased roughly by 10%
  
• 
  AMP does something similar (with signaling), reducing the bit rate by roughly 2% but only increasing encoding time by 33%, with negligible decoder complexity impact.
  
• 
  No action taken.
  

JCTVC-C183 Enhanced switching of interpolation filter for HEVC [T. Yoshino, S. Naito, S. Sakazawa (KDDI)]

Adaptive Interpolation Filter (AIF) technology (e.g., as in prior KTA work) has been proposed to improve the interpolation filter performance for fractional-pel motion compensation. However, this scheme requires high computational cost for calculating the filter coefficients on the encoder side. This contribution proposed an Enhanced Switch Interpolation Filter (ESIF) approach that pre-defines several interpolation filter coefficients in order to reduce the computational cost for calculating the coefficients. This scheme allows updating the number of filter coefficient sets and the filter coefficients slice by slice. The experimental results reportedly showed that the BD-bit rate against AVC reached 1.9% under CS2 condition. Remarks recorded during the discussion included the following:

• 
  The Powerpoint presentation did not match with the contribution (a set of 8 filters was proposed instead of 4).
  
• 
  The current scheme has 500% coder complexity.
  
• 
  The "Low complexity scheme" with 130% coder complexity for filter decision gains 0.3-0.4% on average.
  
• 
  No action taken.
  

JCTVC-C204 Bi-directional optical flow [A. Alshin, E. Alshina (Samsung)] (missing prior, available first day)

A modified B-slice prediction was proposed. By combining the optical flow concept and high accuracy gradients evaluation this proposed technique allows sample-wise refinement of motion. This approach does not require any signaling about motion vector refinement to the decoder.

• 
  Sample-wise refinement is performed by solving optical flow equations between two reference images. For gradient calculation, DCT-based interpolation is used.
  
• 
  Uses an anchor with both MRG and skip/direct and DCT-IF12 used for comparison
  
• 
  The reported gain was roughly 2%, with decoder runtime roughly 150%
  
• 
  It was suggested to include this in an experiment on decoder-side estimation
  

JCTVC-C215 Adaptive MV resolution with directional filters [K. Ugur, J. Lainema (Nokia)]

This contribution presented an adaptive motion vector resolution concept using 1D directional filters. The goal of this contribution was to demonstrate the possible coding efficiency gains in interpolation with low complexity. It also presented a design of adaptive motion vector resolution in TMuC, with directional filters. Remarks recorded included the following:

• 
  Saves computations at decoder through additional 8-pel positions (not a dense grid of all possible positions) which allows more 1D interpolations.
  
• 
  Encoding time increased to 124% (due to additional search), decoding time decreased to 80% (although the time measurement results may not be reliable).
  
• 
  Bit rate reduction was reported at 1.3% on average; where the gain comes due to additional 8^th pel positions; and the gain was mostly in class D.
  
• 
  Bilinear interpolation is used for chroma, whereas the LC version of TMuC may have a rounding problem; and it was suggested that this could explain part of the gain.
  
• 
  It was suggested to study this further in the context of AMVRES and MC interpolation.
  

JCTVC-C251 Overlapped block motion compensation in TMuC [P. Chen, R. Panchal, W.-J. Chien, M. Karczewicz (Qualcomm)]

In this contribution, some results of overlapped block motion compensation (OBMC) in geometry partition coding were reported. The use of OBMC to asymmetric motion partitions was also proposed, with simulation results in support being presented. Remarks included the following:

• 
  Is combined with geometric partitioning
  
• 
  Geometric partitioning reportedly gives 3.9% BR reduction for RA, 3.6% for LD; and 0.8% of this is due to OBMC.
  
• 
  If combined with AMP, the gain of OBMC is 0.4% and 0.3% for RA & LD, respectively.
  
• 
  The increase of complexity at the decoder was asserted to be marginal.
  
• 
  The presentation did not match with the available document, and it was indicated that a new version would be uploaded.
  

JCTVC-C301 Geometry motion partition [P. Chen, L. Guo, M. Karczewicz (Qualcomm)] (late registration, missing prior, available first day)

Geometry motion partition has been integrated into TMuC. The software implementation is based on the design proposed in JCTVC-A121. In the software, geometry motion partition is applied to CU size of 16x16 and 32x32. Geometry mode decision has two stages. First, fast geometry mode decision pre-selects 16 best modes based on SAD calculation. Then the best mode is selected based on full RD. Geometry modes are tested only if the best non-geometry mode is a non-skip mode. Remarks recorded included the following:

• 
  For results, see above under JCTVC-C251
  
• 
  Encoder runtime is 200%; geom. Partition test only done after skip decision.
  
• 
  It was suggested to include this in a core experiment on asymmetric & geometric partitioning (previously in TE3).
  
• 
  Note: checking of the sample membership to the left or right partition may also be of concern for complexity; if masks are used for that purpose, it should be considered that the number is low.
  
• 
  The presentation did not match with the available document, and the contributor was requested to upload a new version.
  

JCTVC-C303 Geometry adaptive block partitioning (GEO) cross-check [E. Francois, P. Bordes (Technicolor)] (late registration, missing prior, available first day)

This cross-check contribution reported exactly the same results in BR reduction as JCTVC-C301 (3 class B sequences were missing). No checking of runtime was reported.

JCTVC-C254 Test result of SIFO for sequences with illumination change [R. Panchal, M. Karczewicz (Qualcomm)]

SIFO was asserted to be not just an interpolation filter but it also to have the capability to compensate for illumination changes. SIFO results were provided for some sequences which had illumination changes to highlight that for such sequences, the offset component of SIFO provides useful gains. The test was done using TMuC 0.8 in the low delay configuration. Remarks recorded included the following:

• 
  The presentation did not match with the available document, and uploading of a new version was requested.
  
• 
  It was asked why should interpolation and illumination compensation be combined? It was suggested that these are two different issues (the latter also applying to full-pel positions). The answer by the proponents was that it would be possible to do it separately.
  

JCTVC-C295 Re-compensation based on partial coefficients [H. Zhu (Zhu)] (late registration, missing prior)

The proposal reportedly provided a new idea for video coding. The algorithm divides the quantized DCT coefficients to two parts, using the first part to get a first reconstruction, and using the first reconstruction R0 and the compensation signals Cmp0 and Cmp1 to compute a new weight for the motion compensation. No experiment results were provided, and no action was taken in response.