International organisation for standardisation organisation internationale de normalisation

5.21Encoder optimization

5.21.1.1.1.1.1.1.1JCTVC-H0178 Encoder speed-up for the MV cost estimation [G. Laroche, T. Poirier, P. Onno (Canon)]

This contribution proposed a non-normative (encoder-only) change regarding the introduction of an early termination rule. This rule deals with the computation of the rate-distortion cost of the motion vector predictors at the encoder side. More precisely, a termination rule was proposed to avoid evaluating all the rate-distortion costs of the merge candidates. It was reported that with this early termination rule, an average decrease of 5% of the encoding time is observed without any impact on the BD BR measure (average BD BR gain of 0%).

The modification was indicated to be 10 lines of modification in the HM code.

It was suggested to expose this as a config file option rather than a macro-enabled feature (and, as a general rule, to have such features be switchable in this fashion).

Decision (SW): Adopt (and enable it in common conditions – but keep in mind that experiments affecting AMVP should perhaps operate with it disabled).

5.21.1.1.1.1.1.1.2JCTVC-H0241 Cross-check report for Canon encoder speed-up for the MV cost estimation (JCTVC-H0178) [S.-C. Lim, H. Y. Kim, J. Lee, J. S. Choi (ETRI)] [late]

5.21.1.1.1.1.1.1.3JCTVC-H0213 Rate control based on unified RQ model for HEVC [H. Choi, J. Nam, J. Yoo, D. Sim (KWU), I.V. Bajić (SFU)]

This contribution presented a rate control scheme based on a "unified RQ" model for HEVC. The proposed rate control works at various levels such as the GOP, frame, and unit levels. The unit can be one LCU or a group of LCUs. To solve a "chicken and egg" dilemma between actual generated bit rate and QP value in rate control problem, this contribution proposed a quadratic pixel-based "unified rate-quantization" (URQ) model for HEVC rate control. The proposed URQ model is defined to show the relationship between a target rate and QP value for a pixel. It is asserted that because the proposed RQ model is designed for a pixel, it can be used for any sizes of a rate control level consisting of multiple pixels. In addition, a MAD factor was employed to predict the texture complexity of a rate control level with a linear model. The target number of bits for the current frame was calculated according to a given frame rate, available channel bandwidth, occupancy of virtual buffer, target buffer level, and initial buffer status. For the unit level, target bits are calculated with the number of pixels of a unit, occupancy of virtual buffer, and target bits for the current frame. For evaluation of the proposed algorithm, HM5.0 which uses hierarchical QP setting, was tested.

The number of lines of code was asserted to be less than 1000, with 99% of this put into a single file.

Both a #define and config file parameter control mechanism were provided.

An apparent error was pointed out in regard to the use of QP versus quantization step size.

Delta QP is used at the LCU level.

Some adjustment is included (without operation of real rate control) for non-reference frames, to avoid wasting too many bits on non-reference frames.

Decision (SW): Adopted (with discretion to fix the QP issue), subject to prioritization with other software integration work (and assuming that the code is reasonably clean).