JVET-C0023 Predictors Elimination Technique for HEVC [M. Korman, O. Prosekov (Synopsis)]

(Presented Tue 13:15.)

The HEVC specification defines multiple different Intra Prediction modes for spatial prediction on the PU level. Indexes identifying selected prediction modes for luma and chroma prediction blocks constituting each CU are coded into the CU header. Prediction mode coding involves an orientation-aware MPM selection technique impacting the predictability of intra prediction directions and thus reduced representation of mode index information. Intra Mode indexes still may occupy up to 50% or even 70% of the rate for Intra-only coded sequences and still images, depending on prediction unit size.

While HEVC intra prediction algorithms are all different, for a particular square region of an image application of those algorithms may lead to predictor blocks that do not differ from each other, or to predictor blocks that can be considered close to each other with respect to a given QP value and thus grouped, or eliminated. Grouping reduces the source alphabet for mode indexes, and the reduced alphabet leads to a shorter code required for prediction algorithm representation in the bitstream. On the other hand, modes grouping might lead to direction information lost and, as a consequence, MPM selection coding efficiency reduced.

This contribution proposes a QP dependent similarity criteria for predictor blocks as well as modifications to existent intra modes coding that allow content-dependent grouping of equal (or similar) predictors (aka a predictors elimination technique) for the price of computational complexity in the order of the prediction modes number. A modified MPM selection algorithm enabling orientation-aware operation on group numbers in place of mode indexes is also proposed.

Parsing of content-dependent group indexes can only be by performed by the decoder when all the pixels required for predictors construction are available, thus the contribution proposes CU bitstream reordering.

Proposed modifications were verified with intra PUs of size 4x4, and average BD-PSNR results computed on the whole range of HEVC quantization parameter values from 1 to 50 reportedly indicate an average of 16% gain in BD-rate and about 0.9 dB improvement in BD-PSNR values on a set of 25 test sequences compared to HEVC operating with the same PU size.

Comparison is made in the context of using fixed block sizes. Results using RD optimization for variable block size decision should be used, this is, de facto, what the proposed “grouping” of 4x4 blocks does in a different way

Investigation of this in the context of the JEM was requested (currently it was tested only for HEVC).

The scheme might be similarly be implemented as encoder-side method to decide about block sizes without making a syntax change.

Further study was encouraged.