International organisation for standardisation

6.20Encoder optimization

See also the section discussing HM settings and common conditions.

6.20.1.1.1.1.1.1.1JCTVC-F045 Early Termination of CU Encoding to Reduce HEVC Complexity [R. H. Gweon, Y.-L. Lee (Sejong Univ.), J. Lim (SK Telecom)]

Comparing to AVC, the computational complexity in HEVC has been increased due to newly added efficient coding tools. One reason of the high complexity of HEVC is that every Prediction Unit such as 2Nx2N, 2NxN, Nx2N and NxN is encoded regardless of the performance of the previous encoded Prediction Unit. By using the signalling information coded_block_flag in the HEVC syntax, the Prediction Unit which could have the best performance can be predicted. Therefore, this contribution introduces a method to reduce encoding complexity of HEVC by simply investigating the coded_block_flag. The encoder complexity of the proposed method reduces encoding time by approximately 42% with Y BD-rate loss of 0.85% compared with that of the HM3.2 encoder.

This was noted to use only 6 lines of code.

The group agreed to adopt this into the reference software (switchable and disabled in the common conditions, with integration priority and timing to be determined in coordination with the software coordinator).

6.20.1.1.1.1.1.1.2JCTVC-F092 Coding tree pruning based CU early termination [K. Choi, E. S. Jang (Hanyang Univ.)] [late upload 07-05]

In this proposal, coding tree pruning based CU early termination is suggested for reducing the encoding time in the HEVC test model software. The experimental results showed that the proposed method achieved about 42% reduction in encoding time compared to the HEVC test model 3.1 encoder with only a negligible loss of luma BD-bitrate (i.e., < 0.6%) and a small gain of chroma BD-bitrate (up to 0.8%).

4 lines of code.

The group agreed to adopt this into the reference software (switchable and disabled in the common conditions, with integration priority and timing to be determined in coordination with the software coordinator).

6.20.1.1.1.1.1.1.3JCTVC-F296 Modifications for CAVLC RDOQ [M. Karczewicz, L. Guo, X. Wang (Qualcomm)] (also see cross-check JCTVC-F723)

In this contribution, two modifications to CAVLC RDOQ are presented. In the first modification, the encoder skips some locations for “last position” checking based on coefficient levels. The second modification aims at improving the bits estimation. The performance of each modification as well as their joint performance was reported. It is reported that combining them together can speed up the encoding process (5% for Intra and 1%-4% for inter) and also slightly improve the B-D rate performance (0.1%).

Cross-check reported in JCTVC-F723, and that contribution reports that the software is implementing what is described in the contributions.

The group agreed to adopt this into the reference software (switchable and enabled in the common conditions, with integration priority and timing to be determined in coordination with the software coordinator).

6.20.1.1.1.1.1.1.4JCTVC-F386 Chroma RD cost computation in HM3.0 [T. K. Tan, F. Bossen (NTT Docomo)]

This contribution observes that the rate-distortion cost computation in the HM does not take into account the difference in the QP values between the luma and chroma components. This results in an inaccurate rate-distortion optimization that penalizes the quality of the chroma component at low bitrates. A rate-distortion cost computation that takes the difference into account was implemented. Simulation results show improvements in the chroma fidelity at the cost of a slight luma BD-rate loss.

The average Y BD-rate loss for all coding conditions is 0.7%. The average U and V BD-rate gains for all coding conditions are 7.0% and 7.2%, respectively.

A weighted YUV BD-rate is also reported, with the average Y, U and V PSNR’s of each sequence being weighted by the ratios 0.75:0.125:0.125, respectively. The average weighted YUV BD-rate gain for all coding conditions is 1.1%.

Adopted – enabled by default (and switchable, of course).