International organisation for standardisation

6.16Transforms

6.16.1Core transform implementation

6.16.1.1.1.1.1.1.1JCTVC-F193 Accuracy improvement of Cisco and TI's transform (JCTVC-E243) [Y. Sugito, A. Ichigaya, S. Sakaida (NHK)]

This contribution proposes "high accuracy" 16x16 and 32x32 transform bases based on HM 3.0's transform proposed by Cisco and TI. The word "accuracy" was asserted to mean orthogonality and isometry (normality of transform) of basis vectors. Proposed bases were reportedly verified according to CE10's coding efficiency. The results for the normal QP range (QP=22, 27, 32, 37) and the high QP range (QP=36, 42, 47, 51) reportedly show around 0% average gains for luma. Almost all of the results of sequences for the low QP range (QP=1, 5, 9, 13) show gains for luma; the peak gain is 1.2% without any increase in process time.

Gain at low QP seems most probably due to having a higher degree of orthonormality. The question was asked whether this was penalized by slightly higher ripple that can be observed in some of the basis functions? (In general, basis functions were designed from the HM version by using deviations of +/-1.)

Remark (JO) about CE (not related to this contribution): Would it be useful to perform a visual comparison, e.g. evaluating different transforms when encoding test charts at high QP?

6.16.1.1.1.1.1.1.2JCTVC-F448 Cross-check of NHK’s contribution on accuracy improvement of the JCTVC-E243 core transforms (JCTVC-F193) [A. Fuldseth (Cisco)]
6.16.1.1.1.1.1.1.3JCTVC-F563 Non-Square Transform for 2NxN and Nx2N Motion Partitions [L. Guo, J. Sole, R. Joshi, P. Chen, X. Wang, M. Karczewicz (Qualcomm)]

The current HM software only employs square-shaped transforms. In this contribution, for inter non-square PUs, i.e., 2NxN PU and Nx2N PU, additional non-square transforms 2NxN and Nx2N are integrated in the RQT structure. Non-square transforms achieve an average BD-rate reduction of 0.2% for the RA HE case, 0.6% for LB HE, and 0.7% for LP HE. For the low complexity, the average BD-rate reductions are 0.1% for RA LC, 0.6% for LB LC and 0.6% for LP LC.

Further study encouraged in new CE2.

6.16.1.1.1.1.1.1.4JCTVC-F624 Crosscheck for Qualcomm's Transform in JCTVC-F563 [Y.-W. Chen, Y.-W. Huang (MediaTek)] [late upload 07-09]

6.16.1.1.1.1.1.1.5JCTVC-F592 Recursive factorization for 16 and 32-point transforms using 4 and 8-point HM 3.0 core transforms [R. Joshi, Y. Reznik, J. Sole, M. Karczewicz (Qualcomm)]

In this contribution, 16 and 32-point transforms are proposed based on the recursive factorization proposed in JCTVC-F352. However the 4 and 8-point transforms in the factorization are replaced by the 4 and 8-point core transforms from HM3.0. This represents a trade-off between the number of multiplications and the number of sequential operations in the transform. As an added benefit, the amount of memory required for storing the quantization and dequantization matrices decreases compared with the full butterfly factorization in JCTVC-F352.

Supplements the proposal in CE10 which is based on full factorization. Compromise sought to reduce the sequential operations.

If used in CE10, only one design should be brought and decided early

6.16.1.1.1.1.1.1.6JCTVC-F595 New results for guaranteeing 16-bit transform dynamic range [K. Misra, A. Segall, L. Kerofsky (Sharp)]

This document reports new results for limiting the dynamic range within an inverse transform. It is asserted that the method guarantees a 16-bit dynamic range limit within the transpose buffer. Additionally, it is asserted that the current HM3.0 design does not guarantee this 16-bit limit. The proposed technique was previously proposed in JCTVC-D071/JCTVC-E411. Here, the technique is adapted to HM3.1 which includes as default the transform proposed in JCTVC-E243. The proposed approach shows a BD-rate impact of 0.0% on coding efficiency for common test conditions. Further evaluation for CE10 configuration settings on HM3.0 demonstrates BD bit rate increase of 0.0%.

With a normally operating the quantizer (and usual input data) this cannot happen, and it actually does not happen in the current test set.

Another possibility would be to enforce by specification that an encoder shall not produce such overruns. This was done in AVC, but it may not be the best solution.

The possibility of overflow is also reduced due to the fact that a right shift was introduced in the quantizer.

Must this be normative? Any decoder implementer could include such clipping e.g. to combat against transmission errors.

Anyway, the problem must be solved either by imposing constraints on the encoder or clipping.

Study in AHG (A. Segall, E. Alshina were suggested as chairs for this) with mandate to work out a solution.

6.16.1.1.1.1.1.1.7JCTVC-F561 Crosscheck of Sharp's proposal on 16 bit dynamic range restriction in inverse transforms [R. Joshi (Qualcomm)] [late upload 07-07]