Organisation internationale de normalisation

9.3CE6b Line/pixel based intra prediction

9.3.1.1.1.1.1.1.1JCTVC-D299 CE6.b Report on Short Distance Intra Prediction Method [X. Cao (Tsinghua), X. Peng (USTC), C. Lai (HiSilicon), Y. Wang (Tsinghua), Y. Lin (HiSilicon), J. Xu (Microsoft), L. Liu (HiSilicon), J. Zheng (HiSilicon), Y. He (Tsinghua), H. Yu (Huawei), F. Wu (Microsoft)]

This document reported experimental results of the short distance intra prediction (SDIP) scheme for core Experiment 6 on intra prediction improvement. SDIP was presented in JCTVC-C101 and JCTVC-C270. By dividing the NxN block into lines or non-square blocks, SDIP can reportedly reduce the energy of the prediction residuals by reducing the distance of predicted pixel and its reference pixels. When integrated into the TMuC 0.9 (HM) software, it reportedly shows 4.3% and 5.7% bit rate saving on average, under all intra high efficiency and low complexity conditions, respectively, with about 50% encoding time increase and no obvious decoding time increase.

A square block which is smaller than 32x32 is divided into several lines or non-square blocks with rectangular shape. In the block, pixels are predicted and reconstructed line by line or rectangle by rectangle. In contrast to previous contribution, UDI can be used within the small blocks. In case of pixel-wide partitions, the same mode is used for 4 adjacent partitions. Transform sizes are also adapted to the block sizes.

The numbers given above relate to a so-called "fast" configuration (encoding time increase is roughly 60% and 45% for HE and LC settings, respectively); decoding time increased by 5%

For LC, RDOQ is turned off in SDIP mode (comment: RDOQ of LCEC is anyway not performing well for intra).

SDIP is used for approximately 33% of the blocks on average.

The following comments were recorded in the discussion:

• 
  This introduces substantial changes to block structures
  
• 
  Adds new transform sizes as well (even though based on the same 1D separable bases) + scanning
  
• 
  Not using RQT for SDIP signalling.
  

9.3.1.1.1.1.1.1.2JCTVC-D124 CE6: Verification results of Huawei/ Hisilicon and Microsoft proposals [Jungsun Kim, Byeongmoon Jeon]

9.3.1.1.1.1.1.1.3JCTVC-D191 CE6: Report and evaluation of new Combined Intra Prediction settings [Marta Mrak, Thomas Davies, David Flynn, Andrea Gabriellini]

This report presents various implementations of Combined Intra Prediction (CIP) in TMuC 0.9, in the context of CE6. The presented approach enables selection of CIP settings of different complexities. A selection of CIP settings was evaluated and the results were presented. The results reportedly show that without noticeable increase of complexity, compared to TMuC 0.9, the average BD BR gain for all intra coded sequences is 0.3% and 0.7% for the high-efficiency and low-complexity configurations, respectively. While the gain is reportedly negligible for lower resolutions, on higher resolutions it is reportedly larger - e.g. the average reported gain for Class A is 1.2% and 2.4% for the high-efficiency and low-complexity configurations, respectively. CIP implementation was discussed and parallelizable implementations were presented.

CIP – Combination of directional prediction and local mean prediction (average of decoded neighbor values); prediction is performed open loop. Predictor geometry is adapted based on the position in the block and on the prediction direction.

Usage of CIP can be restricted to certain block sizes (e.g. for low-complexity settings)

Results: "High complexity mode" 1.5/2.5% BR reduction with 75%/117% encoder runtime increase with HE/LC; for "Low complexity" 0.3/0.7% BR reduction with no/4% encoder runtime increase.

It is noticeable that decoder runtimes are slightly increased in all cases.

Sequential processing of pixels is necessary unlike other intra prediction methods, and there is some irregularity (switching of template geometry) at the pixel-level which may be an explanation.

9.3.1.1.1.1.1.1.4JCTVC-D065 CE6: Verification of BBC's contribution on intra coding [Andrew Segall, Jie Zhao]

9.3.1.1.1.1.1.1.5JCTVC-D204 CE6: Cross-check report on combined intra prediction [K. Chono, K. Senzaki, H. Aoki, J. Tajime, Y. Senda]