International organisation for standardisation organisation internationale de normalisation

5.16.5Sign coding

5.16.5.1.1.1.1.1.1JCTVC-H0227 Non-CE11: Sign Data Hiding without RDOQ [G. Clare, F. Henry, J. Jung, S. Pateux (Orange)]
5.16.5.1.1.1.1.1.2JCTVC-H0481 Multiple Sign Bits Hiding [X. Yu, J. Wang, D. He, G. Martin-Cocher, S. Campbell (RIM)]

The contribution proposed to embed multiple sign bits for TUs larger than 4x4. Specifically, as nonzero coefficients and significance maps for large transform units are encoded in 4x4 coefficient sets, for each 4x4 coefficient set that satisfies certain conditions, the sign bit of the first nonzero coefficient would be skipped from the bit stream and later inferred from the parity of the sum of all nonzero coefficients in that coefficient set. Experimental results for high efficiency settings based on HM5.0 reportedly showed an average BD bit rate change of -0.7% (AI-HE), -0.6% (RA-HE), and -0.7% (LD-HE) with average encoder runtime ratios of 105%, 102%, 102%, and decoder runtime ratios of 99%, 100%, 100%, compared to HM5.0 anchors. For LC settings where RDOQ is off, the encoder could apply a minimum distortion criterion to ensure a match between a sign bit and a corresponding parity. Experimental results reportedly showed an average BD bit rate change of -1.3% (AI-LC), -1.4% (RA-LC), and -1.0% (LD-LC) with average encoder runtime ratios of 105%, 101%, 101%, and decoder runtime ratios of 99%, 100%, 100%, compared to HM5.0 anchors. Subjective viewing of decoded sequences reportedly did not reveal observed artefacts that may be caused by the proposed technique.

The main difference compared to H0224 is that this is based on groups of coefficients – typically one coefficient per group is changed with a probability of 0.5. Better results were reported in the low-complexity configuration. The number of coefficients that is changed could effectively be higher than in H0224.

This affects the parsing process, as the existence of the sign bit can only be determined after the significance map is decoded.

One condition is checked per 4x4 block here, compared to one per TU in the case of H0224. In the worst case, the amount of condition checking would be similar.

Proponents of H0224 would also support this proposal as a consistent extension of theirs.

Madhukar Budagavi investigated the impact on implementation complexity and potential impact on visual quality and reported back as follows: Two sequences were viewed (Kimono and Basketball Drive). For Kimono, flickering was slightly reduced, Basketball Drive looked slightly more noisy. No additional blocking artefacts were observed. From the complexity point of view, the original H0224 would have been of concern, because all sign bits were put to the end of the entire block, whereas it would be OK with H0481 where 4x4 subblocks are processed. Additional operations are finding the first and last significant coefficients (which needs somehow to be done anyway in the significance scan), computing their distance, and an XOR operation for the parity.

By appropriate setting of the threshold value, an encoder can decide to not use the method.

One expert raised a concern about throughput, but the method is apparently slightly decreasing throughput.

Decision: Adopt H0481.

5.16.5.1.1.1.1.1.3JCTVC-H0503 Cross-check of RIM multiple sign bits hiding (JCTVC-H0481) [J. Sole (Qualcomm)] [late]