Of itu-t sg16 wp3 and iso/iec jtc1/SC29/WG11

5.3RCE3: Intra prediction techniques (13)

5.3.1RCE3 summary and general discussion

(Reviewed Thu 24th morning GS & JRO)

JCTVC-O0037 RCE3: Summary report of HEVC Range Extensions Core Experiment 3 on Intra Prediction techniques [A. Saxena, D. Kwon, M. Naccari, C. Pang]

This is a summary report on HEVC Range Extensions Core Experiment 3 on intra prediction techniques. The core experiment investigated sample adaptive prediction for various oblique modes, and a nearest neighbor interpolation technique which replaces bilinear interpolation for oblique intra prediction modes. Performance of the proposed methods as well as their combinations was evaluated for both lossy and lossless configurations based on the test conditions and sequences described in JCTVC-N1123.

• 
  Tool A.1 JCTVC-O0080, RCE3: Results of Test A.1 on sample based intra prediction for lossless coding, J. Zhu, W. Zheng, K. Kazui (Fujitsu)
  
  • 
    A.1.1 Lossless some gain on SCC 4:4:4, loss on "RExt" camera content – some results not especially encouraging
    
  • 
    A.1.2 Lossy and lossless – some results not especially encouraging
    
• 
  Tool A.2 JCTVC-O0047, RCE 3: On sample adaptive intra prediction for oblique modes in lossless coding H. Chen, A. Saxena, F. Fernandes (Samsung) and JCTVC-O0048, RCE 3: On sample adaptive intra prediction for oblique modes in lossy coding, A. Saxena, H. Chen, F. Fernandes (Samsung) – some gain, but adding extra modes and gain is not so much
  
• 
  Tool A.3: JCTVC-O0049, RCE 3: Nearest-neighbor intra prediction for screen content video coding, H. Chen, A. Saxena, F. Fernandes (Samsung) – some gain (esp. 4:4:4 and SCC – roughly 1.6% for YUV and 2% for RGB)
  
• 
  Subtest B, Tool B.1: JCTVC-O0051, RCE 3: Combination of sample adaptive prediction and nearest neighbor prediction for oblique modes, A. Saxena, H. Chen, F. Fernandes (Samsung)
  

A non-CE contribution has two variants: 1) implicit derivation of the flag, or 2) coupling transform skip to nearest-neighbour interpolation.

The gain did not seem sufficiently compelling.

Overlapping notes below.

A1 and A2 are using pixel-wise DPCM (with samples used for prediction dependent on directional modes, different variants in which of the directional it is used

A3 is using the same directional prediction as HEVC, but it can be signalled that the nearest boundary sample (instead of an interpolated position) is used as prediction, i.e. switching off the interpolation filter

A1 generally moderate gains for screen content in lossless coding, moderate losses in lossy coding

A2 similar in lossless, but also very small gain (no losses) in lossless coding

Both A1 and A2 include additional modes, i.e. add implementation complexity

A3 has moderate gain for lossless coding (usually 0-0.2% except 1% for 4:4:4 screen content)

For lossy coding, it has 1.6/1.4/0.9% gain (AI/RA/LD) on 4:4:4 YCbCr screen content in main tier.

(without the “easy to code” screen sequences that were removed from the CTC test set)

Increases the encoder complexity (requires additional checking of modes that switch off interpolation), and also the decoder needs to implement more circuitry.

No action on any of the proposals

No action.