JVET-B0059 TU-level non-separable secondary transform [X. Zhao, A. Said, V. Seregin, M. Karczewicz, J. Chen, R. Joshi (Qualcomm)] [late]
Chaired by J. Boyce.
In this contribution, a TU-level non-separable secondary transform (NSST) is proposed. Compared to the current CU-level NSST design in JEM-1.0 software, the proposed method speeds up the encoder by reducing the number of rate-distortion checks and using transform-domain distortion estimation. With the proposed TU-level NSST, an average 44% overall encoder run-time reduction is reportedly achieved over JEM-1.0 for all intra (AI), and average BD-rate improvement of 0.1% is reportedly achieved for the luminance component for both AI and random access (RA) configurations.
Replacing CU-level CSST with TU-level NSST provides an encoder speedup of 44%, plus a small coding gain in luma and coding loss in chroma.
A new Hypercube-Givens Transform (HyGT) is used in the computation of the secondary transform and also proposed. It has a butterfly-like structure. Separate experimental results are not given for the two aspects, which would be preferable.
A small increase in decoder runtime was seen, but it is unclear why.
Further study was encouraged, and it was requested to consider the interaction with B0051.
This was further discussed on Wednesday.
Decision (SW): Make SW available in experimental SVN branch of JEM software. Include in the EE document for testing. Proponents were asked to provide separate experimental results for each aspect individually.