International organisation for standardisation

General conclusion on quantization:

Continue CE4 on previous subtests 2 & 3, and quant matrices as new subtest.

A general problem in CE4 is still the realistic scenario of test conditions (granularity and amount of QP changes, realistic quant. matrices) – to be further discussed in the preparation of CE description – a BoG was created to discuss this further. See notes elsewhere.

6.12Alternative coding modes

6.12.1.1.1.1.1.1.1JCTVC-F150 Inter modes for screen content coding [W. Zhu, W. Ding, Y. Shi, B. Yin (Beijing Univ. Tech.)]

This contribution introduces two inter coding tools (InterBCIM and InterRSQ) for screen content coding. The two inter coding tools are implemented into the HM2.0 software. The experimental results reportedly show that the bit rate saving tested on screen content is 20% or 3.36 dB improvement on average.

Inter RSQ is, in principle, transform bypass and PCM. This is said to be useful for high rates.

Only 3 sequences were tested.

6.12.1.1.1.1.1.1.2JCTVC-F200 Improvements of the BCIM modes in screen content coding [C. Lan (Xidian Univ.), J. Xu, G. J. Sullivan, F. Wu (Microsoft)]

This document presents improvements of BCIM (Base Color and Index Map) mode to compress screen contents. It further improves the coding efficiency for screen content. The improved mode is an alternative method to provide an upper bound for the coded bits, avoiding unexpected large bit number generation which is prohibitively greater than that of raw data. Experimental results of implementing BCIM in HM3.0 for screen content sequences under the high efficiency intra coding configuration are reported in this document. For the typical screen content sequences, the proposed BCIM mode reportedly achieves 38% bit saving on average. For all the tested screen sequences, a 21.6% bit saving on average was reported. Experimental results over the synthesized noise sequence demonstrate that BCIM can provide an upper bund for the output bits.

BD rate gain is high for specific screen content. For natural video, the gain is low (3% or less).

Only intra was tested.

The results reported only show the luma bitrate (except table 1).

Comments included the following: Is PSNR the right measure? May be misleading, as tools like this reproduce flat areas with zero deviation, whereas transform coding may introduce noise which may be acceptable. Edge ringing could be more severe on the other hand. Subjective comparison?

6.12.1.1.1.1.1.1.3JCTVC-F564 Near Lossless Coding for Screen Content [W. Gao, G. Cook, M. Yang, H. Yu (Huawei)]

This contribution proposes a near lossless coding method that involves modifications to the existing intra coding tools. This method is designed for encoding texts and graphics at high quality and high coding efficiency, and screen sharing is one of potential applications of this method, where computer display signal is treated and processed and delivered as video. Typical screen capture video reportedly consists of a mix of cameral-captured video, computer generated graphics, texts, etc., and fidelity or the quality of the text and graphics are extremely important. The major problems with today’s solution reportedly include poor compression quality, low frame rate, and long latency, and poor quality of text and reportedly can cause eye-fatigue. In summary, it was asserted that neither the compression ratio nor compression quality with the existing solutions are satisfactory. To address this asserted problem, a set of modifications to the existing tools were proposed to achieve both near-lossless quality and high compression efficiency in coding of screen compound video. For purpose of evaluation, these modifications and the associated syntax element changes have been integrated into HM3.0-dev-SDIP branch. Two test sequences proposed to the SCC Ad-Hoc group are used in the tests. Compared to baseline HM3.0-dev-SDIP software, the proposed method can reportedly maintain the high coding efficiency of HM while achieving near lossless quality for the text regions.

SDIP is reported to give 13% rate benefit (intra coding) for two screen content sequences, with various tools (ALF, deblocking filter, …) being turned off.
6.12.1.1.1.1.1.1.4Conclusions from this:

It is interesting to see that tools exist that formally give high gain for this specific content.

Transform bypass could be interesting – this leaves CABAC as plain spatial context-based AE.

Usage of PSNR should be investigated.