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

JCTVC-O0233 Motion Compensation Interpolation with 16 Bit Intermediate Buffer for HEVC Range Extension [W. Pu, J. Chen, W.-S. Kim, M. Karczewicz, J. Sole, L. Guo (Qualcomm)]

(Reviewed Fri 25th morning GJS & JRO)

In current HEVC Range Extension design, the intermediate buffer of motion compensation interpolation module reaches up to 20 bits to retain high intermediate data accuracy. In this proposal, a method is proposed to keep the motion interpolation module’s internal data within 16 bit.

The worst case is half-pel position { -1, 4, -11, 40, 40, -11, 4, -1 }.

Current draft:

• 
  1st stage
  
• 
  shift1 = Min( 4, BitDepth − 8 )
  

• 
  8 b  16 b (no right shift)
  
• 
  10 b  16 b (2 bits right shift)
  
• 
  12 b  16 b (4 bits right shift)
  
• 
  14 b  18 b (4 bits right shift)
  
• 
  16 b  20 b (4 bits right shift)
  
• 
  2nd stage has right shift by 6, no clip (prior to WP & bipred combination)
  

Proposal:

• 
  1st stage
  
  • 
    > 12 b introduce clipping to 16 b [0, 65535]
    
  • 
    > 14 b increase shift1 to Max( 4, BitDepth – 10)
    
• 
  2nd stage introduce clipping to 16 b [0, 65535] (prior to WP & bipred combination)
  

The sequences tested were SVT sequences – there was some concern about the quality of the LSBs in these sequences.

It was remarked that the high-precision forward transform was not used here, and tests using that would be desirable.

It was remarked that coding performance for higher bit rates and lossless would be good to study.

It was remarked that there is an offset trick in the current software that saves 1 b of dynamic range and that this trick could be used at high bit depths as well to reduce the necessary right shifting by 1 b.

It was remarked that the coding results show no difference from the current method at 14 b, so a simpler approach just shifting by more bits might suffice for that.

A proponent indicated that this was mostly due to the quality of the test material – more difference would be seen with higher precision test material.

It was remarked that this effectively results in having no precision beyond that of the input data when the input data is 16 b.

Clipping basically can save 2 bits when applied.

It was noted that the transform stage goes well beyond 16 bits for its processing, so it may not be necessary to worry so much about the dynamic range at the MC interpolation stage.

No action.