International organisation for standardisation organisation internationale de normalisation

CE4: Illumination compensation

1. Summary

The most critical complexity comes with adaptive shifts (ShiftA).

Second critical operation is get MSB (irregular).

From that point of view, the most promising candidates are Test 1.3 and E0046.

Test 1.3 has 0 gain/loss, E0046 has 0.2% BR reduction.

Test 1.3 has 169/0/0 Mul/Add/shiftC for 8x8 which is the most critical CU size.

E0046 has 107/85/67 Mul/Add/shiftC for 8x8.

Several experts express opinion that 0.2% coding gain is worth to be paid by approx 1 mul/sample.

Cross-checker confirms that WD text is matching the software.

Decision: Adopt: JCT3V-E0046.

Test 2: Aligned operation for Bi-prediction

Test 2 consists of aligning the operations between the uni-prediction and bi-prediction when IC is enabled.

In decoder side, the alignment enables uni-prediction compensation (identical motion compensation) if L0 motion parameter and L1 motion parameter is identical in IC case.

Note: Current HTM’s identical motion compensation is buggy because it doesn’t consider if IC on or off. But this bug doesn’t appear because encoder and decoder have the same bugs so there is no enc/dec mismatch.

In encoder side, the alignment is beneficial because L0 block and L1 block can be reused for Bi block in motion estimation stage.

The current specification of IC defines different clipping for cases of uni and bi-prediction. JCT3V-E0168 investigates using the same clipping (from uni prediction) for both cases. Therefore bi-prediction where identical MV are in L0 and L1 can be replaced by a uni-prediction which would not be possible otherwise.

Decision: Adopt JCT3V-E0168.

2. CE contributions

3. Related contributions

Decision: Adopt (see under summary above).

15.0.1.1.1.1.1.1.118JCT3V-E0135 CE4.h related: Coding of illumination compensation flag [J. Kang, Y. Chen, M. Karczewicz (Qualcomm)]

In the current 3D-HEVC, a CU-level flag for illumination compensation (IC) (i.e., ic_flag) may or may not be signaled in a different condition depending on a slice-level flag and a prediction mode without parsing dependency problem. This proposal presents modifications on the ic_flag signaling including the following aspects. First, a slice-level flag is removed and replaced by checking whether the current slice belongs to an IRAP picture. Second the derivation of the ic_flag signaling condition is simplified by skipping to check the reference picture indices due to disabling AMVP in the IC mode. Third, the CABAC context of the ic_flag is modified. Experimental results show that the two modifications provide the only minor change (0.1% gain) in coding gain in the common test condition (CTC).

Two modifications are suggested:

• 
  Replace slice_ic_disable_merge_zero_idx by an implicit derivation from NAL unit type
  
• 
  Modify CABAC contexts for ic_flag (more contexts)
  

The second modification provides 0.1% BR reduction, and is similarly proposed in JCT3V-E0192 (see disposition there).

15.0.1.1.1.1.1.1.119JCT3V-E0143 3D-CE4.h related: Illumination compensation for BVSP and adaptive IC model [M. W. Park, J. Y. Lee, H.-C. Wey, C. Kim (Samsung)]

In the current 3D-HEVC, the ic_flag is signalled for every BVSP mode to remove parsing dependency, however, the BVSP mode does not apply the illumination compensation. Therefore, it is proposed to simply apply the illumination compensation to BVSP mode with the offset model. This proposed method reportedly provides 0.1% bit-saving on coded and synthesized views. In addition, it is observed that if IC model adaptively can apply to each component (i.e. luma, chroma, and depth map), more coding gain could be obtained and the complexity also could be decreased. When the offset model is applied to chroma, while other components use the current IC model, the proposed method reportedly provides 0.1% bit-saving on coded and synthesized views. Additionally, when the offset model is applied to depth, while other components use the current IC model, the proposed method reportedly provides no coding loss. When combining all propose methods, 0.2% and 0.1% bit-savings are reportedly obtained for coded and synthesized views, respectively.

Illumination compensation suggested here is different from the current one (which does not provide gain with BVSP).

In terms of maximum complexity, useful to have it disabled.

Additional complexity not justified by 0.1% gain.

15.0.1.1.1.1.1.1.120JCT3V-E0192 CE4.h related: Context model for illumination compensation flag [J. W. Jung, J. Heo, S. Yoo, S. Yea (LGE)]

In 3D-HEVC, illumination Compensation(IC) process is specified by IC CU flag. The conventional IC flag has been encoded by only one context model. In this contribution, the context model is separated to three context models derived from neighbour block. It has BD-rate change -0.10% compared to HTM-7.0r1 on synth-total result with 100.5% encoding time, and 99.8% decoding time.

Slightly different definition of context models than in E0135; similar small gain

Both proposals achieve this gain by making the ic flag coding dependent on the neighbor blocks (i.e. increasing complexity of entropy coding

No support for adoption or further investigation by other experts.

No action.

15.0.1.1.1.1.1.1.121JCT3V-E0196 CE4.h related: Cross-check on resampling in IC parameter derivation and 4x4 Chroma removal(JCT3V-E0046) [J. W. Jung, S. Yea (LGE)] [late]

5. 
   
   Directory: sc29 -> open -> 29view
   29view -> International organisation for standardisation organisation internationale de normalisation
   29view -> International organisation for standardisation organisation internationale de normalisation
   
   
   
   Download 8.47 Mb.
   
   Share with your friends: