Abstract not available in document
JVET-D0031 Residual Coefficient Sign Prediction [F. Henry, G. Clare (Orange)]
Discussed Sunday 1500 (GJS & JRO)
This contribution proposes a method of prediction for the signs of luma residual coefficients. A number of signs per TU can be predicted, limited by a configuration parameter and the number of coefficients present. When predicting n signs in a TU, the encoder and decoder perform n+1 partial inverse transformations and 2n border reconstructions corresponding to the 2n sign combination hypotheses, with a border-cost measure for each. These costs are examined to determine sign prediction values, and the encoder transmits a sign residual for each predicted sign indicating whether the prediction for that sign is correct or not using two additional CABAC contexts. The decoder reads these sign residuals and later uses them during reconstruction to determine the correct signs to apply after making its own predictions.
It was asserted that this method of sign prediction, when predicting up to 4 signs per TU, brings Y gain improvements of 0.5% (AI), 0.6% (RA), 0.5% (LD), 0.5% (LP), with associated complexity measures of 145%, 123%, 118%, 124% for the encoder, and 142%, 103%, 105%, 109% for the decoder.
The decoder needs to have the reconstructed coefficients available before parsing the signs.
A participant indicated that JCTVC-A115 had a somewhat similar proposal.
Some interest was expressed to further investigate this, but the encoder and decoder complexity should be further reduced.
Further investigation in an EE was planned.