Organisation internationale de normalisation

10.3ROT improvements

10.3.1.1.1.1.1.1.1JCTVC-D357 (m19125) CE7: Experimental results of fast ROT by Samsung [E. Alshina, A. Alshin, F.-C. Fernandes, Y. Piao, W.-J. Han (Samsung)]

This contribution provides simulation results when Rotational transform (ROT) is enabled in TMuC 0.9 against the TMuC 0.9 anchor results. In Intra-Only test case, about 2% (HE) / 3% (LC) gain was reportedly provided by ROT. In Random-Access test scenario ROT, gain is over 1% on average. No additional changes (like scan modification, for example) typically improving alternative transform performance were reportedly applied in this test.

Number of rotation matrices to be stored – horizontal/vertical 4 rotation variants for 4x4 and 8x8 each.

Integer rounding is applied, but in a manner such that orthogonality may be lost.

Implementation of ROT by matrix multiplication, due to sparsity of the set of coefficients on which the transform is applied, is claimed to require a lower number of operations than with an alternative (e.g. DST) transform. A lifting implementation was (verbally) reported to exist.

BR reduction 1.9%/2.9% in HE/LC intra only.

With "fast" algorithm, encoder runtime roughly 160%/200%; decoder runtime 97%/105%.

Restricted ROT with only best UDI mode: BR reduction 1.1%/1.9%, encoder roughly 120%.

Note: Loss of performance in chroma components.

10.3.1.1.1.1.1.1.2JCTVC-D030 (m18773) CE7: Cross-verification of Samsung's (JCTVC-D357) Fast Rotational Transform [R. Cohen, A. Vetro, H. Sun (Mitsubishi)]

10.3.1.1.1.1.1.1.3JCTVC-D409 (m19201) CE7: A crosscheck of Samsung’s fast ROT proposal (JCTVC-D357) by Qualcomm [P. Chen, R. Joshi]