Of itu-t sg16 wp3 and iso/iec jtc1/SC29/WG11

6.1.6Lossless and screen content coding related contributions

See also section 3.6 on HEVC and RExt use cases.

JCTVC-M0304 Non-RCE1/Non-RCE2: Simplified level coding for transform skip and lossless coding [S. H. Kim, A. Segall (Sharp)]

In this document, a modified level coding for a transform skip block is presented. Except for removing greather_than_1 coding and greather_than_2 coding, there is no change to the level coding in the proposed modified level coding. The modified level coding is proposed to be employed/activated when the transform skip flag is equal to 1. Simulation results reportedly show that the average BD-rates of the proposed method for the three color components are 0.0%, 0.0%, and 0.0% for AI-Main-tier, 0.0%, 0.0%, and −0.1% for RA-Main-tier, and 0.0%, 0.1%, and 0.0% for LD-Main-tier for ordinary common conditions without using screen content test sequences.

The modified level coding method is proposed to always be activated in lossless coding because transform and quantization process is not used. Hence, worst case context coded bins per sample has been reduced from 1.6 bins/sample (which has been mentioned/discussed in JCTVC-H0728) to 1 bin/sample. It is reported that the overall bit rate increase of the modified level coding is 0.8% for AI, 1.8% for RA, and 1.8% for LD, respectively, for lossless coding for the RCE2 tested sequences. To improve lossless coding performance, remaining coefficient coding process has been additionally modified on top of the modified level coding by signalling the Rice-parameter for each sub-block level and removing the Rice-parameter update process. In terms of computational complexity, worst case context coded bins per sample for transform skip blocks has reportedly been reduced from 1.6 bins/sample to 1.1 bin/sample, because 1 context coded bin is added for rice parameter coding. With the reduction of the worst-case context coded bins for transform skip blocks, simulation results reportedly show average BD-rates change for the proposed method are −3.5% for Al, −1.7% for RA, and −1.3% for LD for lossless screen content coding for the RCE2 tested sequences. However, this benefit is concentrated in SC GBR sequences – there are significant losses for the Class F and range extensions sequences.

The simplification part of the proposal to change the entropy coder would introduce a compatibility problem w.r.t. the ordinary HEVC entropy coding. At least for a generic profile (vs. some profile specifically made only for non-generic application usage), we would want decoders to be able to decode version 1 bitstreams, and this type of alteration would actually be burden in such a scenario, as it would require support of two entropy decoding modes. We do not currently have a mandate to develop a lossless coding profile.

However, it was planned to test just the Rice parameter part of the proposal in a CE along with other proposals that modify Rice parameter handling.