International organisation for standardisation

5.3.3Discussion and Conclusions

5.4CE4: Quantization

5.4.1Summary

5.4.1.1.1.1.1.1.1JCTVC-F024 CE4: Summary report of Core Experiment on quantization [K. Sato, M. Budagavi, M. Coban, H. H. Aoki, X. X. Li (CE coordinators)] [upload 07-12]

Subtest 1: Signalling MaxCUDepthDQP at LCU level: No gain – further study.

Subtest 2: QP prediction from neighboring CU QP values (9 contributions).

Subtest 3: Adaptive de-quantization offset.

For subtest 1 and 2, a TM5-like rate control is used, mainly using variance of CUs as a criterion to determine the QP (intended to be subjectively adapted). This may not be fully realistic in real-world applications. One expert commented that sometimes this simple approach even produces worse quality.

For subtest 2, a combination proposed inJCTVC-F661 (comb. of 2.3.g&f&e) gives roughly 0.35% bitrate reduction on average and is suggested for adoption. One expert mentions that 2.3.e is not in the original CE description and is in principle a new technique. Another suggested combination (JCTVC-F705) is made from 2.3g&(b||c).

Compared to constant-QP settings, the variable-QP scenario tested here is 6% worse.

Subtest 3: 2 methods (ARL making adaptation based on data, AQO based on Laplacian model). AQO only uses one offset value, but different for each components, ARL uses multiple (variable) offset values.

It seemed not to be clear whether the gain is by re-allocation of bits. With RDOQ on, gain by ARL is 1.1% BR reduction maximum, 0.8% for AQO with lambda refinement. Lambda is adapted on a sequence-by-sequence basis. With RDOQ off, gain is 1.8% and 1.9%.JCTVC-F610 is another similar contribution with only one offset for all three components, and is reported to give even better gain.

Need to identify methods to clarify where the gain comes from. Further study on subtest 3 (CE cont.)

Following these suggestions, JCTVC-F661 and JCTVC-F705 were presented in more detail.