JVET-AQ0172 AHG11: VLOP4 Modification for Complexity Reduction [H. Cho, J. Kim, S. Bahk, T. Lee, H. Y. Kim (KHU), D. Kim, S.-C. Lim (ETRI)] [late]

This contribution modifies VLOP4 by using coding parameter-guided feature modulation to reduce computational complexity. Specifically, this contribution proposes to process the coding parameters among the VLOP4 inputs, namely QPbase, QPblock, and IPB information, through a separate operation and use them to scale and shift intermediate features. The proposed method reduces the complexity while maintaining the Y BD-rate performance. The proposed method has been implemented based on VLOP4 of NNVC-16 using SADL integer operations.

• Complexity
  • Parameter counts: 69.2k (VLOP4 Anchor: 72.72k)
  • MACs/pixel: 4.84k (VLOP4 Anchor: 5.09k)
• Over VLOP4
  • RA: {0.01%, 2.05%, 1.71%}
  • AI: {-0.04%, 1.44%, 1.94%}

Additionally, the Random-Access performance results for the CfP contents are as follows.

• Over VLOP4
  • RA: {-0.05%, 3.2%, 1.54%}

It was asked whether the scaling and shift used in the CPFM block is necessary in terms of compression benefit.

It was asked if the large drop in chroma is due to replacement of 3x3 convolution by 1x1? Has not been investigated – proponent reports that they are currently running an additional test with other luma/chroma balance in training.

It was commented that the benefit of using local QP as input might be rather low in CTC where constant QP is used.

Only small interest was expressed for investigation in EE at this moment. Further study recommended along the questions above.