Joint Collaborative Team on Video Coding (jct-vc)

5.20Non-normative: Encoder optimization, decoder speed improvement, post filtering, loss concealment

JCTVC-I0142 Multiplication free interpolation filter implementation [B. Li, H. Li (USTC), H. Yang (Huawei)]

(Reviewed in Track A.)

This contribution proposes a multiplication-free interpolation filter implementation. It is only an optimization at the code level without any RD performance difference. With this multiplication free implementation, about 10% encoding time savings and 8% decoding time savings was reported to have been obtained.

This would be a non-normative implementation optimization (with no text change needed). It was suggested for this to be discussed with the SW coordinator to determin whether this is desirable to include in the reference software. One expert commented that it may sometimes be better if the reference software is not deviating too much from the text description.

The software coordinator thought it would be cleaner to not have this in the codebase.

JCTVC-I0055 Picture quality evaluation of non-local means post filtering by SSIM [M. Matsumura, S. Takamura, A. Shimizu (NTT)]

This contribution reports an asserted benefit of a non-local means post filtering in terms of a picture quality evaluation using the SSIM criterion. Especially, SSIM reportedly improves up to 0.0043 in a low bit rate test condition (QP: 37). Average BD-BR improvements of 1.1%, calculated by SSIM instead of PSNR, were reported, and the improvement of subjective quality in various regions was reportedly confirmed. It was suggested that SSIM-based RDO might be useful for improving the subjective quality. The maximum reportedly gain was 5.1% for the sequence "SlideEditing".

This contribution was just considered for information at this time. The filter strength is selected by the encoder, and a post filter hint would be necessary as an SEI message.

JCTVC-I0094 Improvement of the rate control based on pixel-based URQ model for HEVC [H. Choi, J. Nam, J. Yoo, D. Sim (KWU), I. V. Bajić (SFU)]

This contribution presents a refinement of a rate control algorithm based on the rate control proposed in JCTVC-H0213. This document includes two additional conditions on the previous rate control implementation. One is an initial QP setting for the beginning of a sequence, the other is a boundary for target bits in the frame level rate control. The initial QP is automatically calculated based on bits per pixel (bpp) with target bit rate. The QP value in the configuration file is ignored. The boundary for target bits of a frame is used to conform with the hypothetical reference decoder (HRD) requirement, so the target bits are clipped by lower and upper bounds. In addition, the rate control in JCTVC-H0213 is modified according to experts' comments in the preceding meeting in San Jose, especially using quantization step (Q-Step) in rate-quantization model (RQ model) instead of directly using QP value. The refinement of the rate control is implemented on HM6.1 and evaluated on diverse conditions.

Compared to the current RC algorithm in HM, the BD bit rate loss is reduced from 0.83% to 0.74%. It is not clear whether this reduces encoder run time or not.

No action was taken on this.

JCTVC-I0426 QP determination by lambda value [B. Li, D. Zhang, H. Li (USTC), J. Xu (Microsoft)]

This contribution presents a QP determination algorithm in HEVC. To improve the coding efficiency multiple-QP optimization (including slice level, LCU level, and sub-LCU level) can be applied in HEVC. In the multiple-QP optimization, the lambda value is fixed and the QP value that provides the smallest RD cost is selected as the best QP. But this multiple-QP optimization increases the encoding complexity significantly. If 3 QPs are used in the RDO process, roughly 300% encoding time is required. In the document, QP is determined by the lambda value, according to a pre-trained QP–lambda relationship. When applying the proposed method to HM-6.0, about 1.8%, 1.7%, 1.4%, and 1.4% luma bit rate savings is reportedly obtained for RA-Main, RA-HE10, LD-Main, and LD-HE10, respectively. The bit rate savings on U and V are reported to be somewhat larger than that on Y. The method in this contribution only includes non-normative modifications.

Training was done for classes C and D, but the reported gain also applies to the other classes.

The change of QP over the different levels is more aggressive – using QP+2 or QP+3 instead of QP+1. This may be one of the main reasons for the BD bit rate saving, as the PSNR fluctuations become larger and due to the computation of mean frame PSNRs larger fluctuation gives better mean values. The impact on visual quality was not reported.

Keeping same QP and training the lambda would be more desirable, but that would be difficult for the current HM. In fact, it is said that currently loss was observed for intra coding in this case.

The visual quality impact was unknown.

Further study was encouraged to provide a better understanding of the potential for benefit in this.

The proponent said they would provide software in an update of the contribution.

JCTVC-I0573 Cross-check of QP Determination by Lambda Value (JCTVC-I0426) [Y. Chiu, W. Zhang (Intel)] [late]
JCTVC-I0433 Adaptive Rate Control for HEVC [J. Si, S. Ma, W. Gao (Peking Univ.), M. Yang (Huawei)]

This contribution describes a frame level rate control scheme for HEVC, which originates from an ABR (adaptive bit rate) scheme used in the x264 codec with some fine tuning for HEVC. The proposed rate control scheme is designed specifically for the low delay and random access coding cases, respectively, and implemented into HM6.1. Compared with the original rate control scheme proposed by JCTVC-H0213, the average BD-BR improvement computed using piece-wise cubic interpolation can reportedly be up to 23.7% for RA-main (for LP-main: 18.3%; LB-main: 19.1%).

It was asked whether larger bit rate fluctuations occurred than with current HM rate control scheme.

It was commented that BD bit rate is not a valuable measure in this case – that it makes most sense in the case of constant QP operation.

CPB occupancy variation in comparison to current algorithm would be desirable to study.

Visual quality is more important to judge rate control algorithms than the BD measure.

Further study was recommended.