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

JCTVC-M0086 AHG-17: complexity and performance analysis of SHM1.0 compare to HM8.1 simulcast [A. Alshin, E. Alshina (Samsung)]

This contribution contains performance and complexity analysis of SHM1.0 both IntraBL and RefIdx frameworks, compared to single layer coding. Complexity assessment methodology developed for SCE3 &4 by AhG-17 was used. It is reported that memory access SHM1.0 in the worst case doesn’t exceed HEVC. In average for motion compensation test scenarios (RA and LD-P) the coding efficiency compare to HM8.1 simulcast is 18.5% for IntraBL framework and 18.0% for RefIdx framework in terms of Luma BD-rate. Average memory of SHM1.0 is lower than HM8.1 simulcast: 94–96% for IntraBL and 93–94% for RefIdx frameworks. In average for motion compensation test scenarios SHM1.0 IntraBL framework outperforms RefIdx by 0.7% with cost of 1–2% extra memory access in average.

Presentation to be uploaded.

The contribution conclusions were reported as follows:

• 
  The average computational complexity and memory access of SHM1.0 is reportedly 4–6% lower than for HM8.1 simulcast single layer coding, while 18% Luma BD-rate reduction is achieved in motion compensation test scenarios.
  
• 
  It was proposed that the motion vector in inter-layer predicted blocks in the RefIdx framework shall be normatively restricted to be 0, which was noted to already be a planned constraint. This restriction was asserted to guarantee that the worst case memory bandwidth of SHVC will not exceed HEVC limit.
  
• 
  Complexity assessment using the AhG17 tool shows negligible difference between IntraBL and RefIdx frameworks (1–2% in motion compensation test scenarios), while the IntraBL branch outperforms the RefIdx framework 0.7% in terms of Luma BD-rate (0.6% in terms of Chroma BD BR).
  

The reported test data did not include LB case testing.

All results given here are reporting average results (not worst case).

Complexity comparison in this contribution is done against simulcast; in the discussion it is pointed out that from an application perspective, it would be more reasonable to compare vs. single layer corresponding to enhancement layer, as usually only the higher resolution would be decoded.

RefIDx was investigated once by assuming block-based processing (i.e. performing upsampling only in cases where it is needed), in which case there is no significant difference compared to textureBL. Results are also reported for the case when picture-based processing would be used (upsampling would always be necessary). In that case, the memory bandwidth is higher.

As analyzed, the use of whole-picture upsampling was not considered for the RefIdx approach, as this type of operation has higher memory bandwidth.

However, it was remarked that whole-picture processing can be used as an architecturally/conceptually simple way to construct an SHVC decoder when starting with a single-layer HEVC implementation.

It was noted that the complexity anchor for some comparisons was simulcast decoding of both simulcast streams, which does not necessarily really make sense.