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

JCTVC-M0342 Non-SCE4 Cross-check for 6-Tap Adaptive Resampling Filter (JCTVC-M0271) [E. Alshina] [late]
JCTVC-M0273 Non-SCE4: Switchable Filter on Integer Position [W. Pu, V. Seregin, J. Chen, X. Li, M. Karczewicz (Qualcomm), E. Alshina, A. Alshin, Y. Cho (Samsung)]

The integer position is not filtered in the up-sampling process of the current SHVC test model. This proposal provides simulation results for filtering integer positions in SHVC. Integer position filter coefficients are fixed. But filter is switchable. Two methods of switching are evaluated. In the first one, the encoder selectively enables the fixed smoothing filter for each picture and signals the selection using one bit in slice header. This scheme achieves average BD-rate reduction of 0.13% (spatial) and 1.38% (SNR), respectively, for TextureRL framework. For RefIdx framework, average BD-rate reduction is 0.13% (spatial), 1.67% (SNR), respectively. The second scheme applies to SNR scalability RefIdx framework only, the reconstructed interlayer reference picture and the filtered interlayer reference picture are both inserted into the enhancement reference picture list. This scheme achieves average BD rate reduction of 2.1%.

Relates to SCE 4.1.1. In the RefIdx approach, two reference indexes are assigned to filtered and unfiltered copies of the BL picture so that the filter on/off switch can be applied in the RefIdx approach.

Picture-level on/off was also tested (which is an appealing variation for the RefIdx approach).

	All-Intra		MC		SNR
	Luma	Chroma	Luma	Chroma	Luma	Chroma
TextureRL
M0087 (CU on/off)	NA	NA	NA	NA	−2.1%	0.0%
M0273 (Pic on/off)	−0.1%	−0.0%	−0.2%	0.0%	−1.4%	−0.4%
RefIdx
M0087 (no switch)	NA	NA	NA	NA	−1.4%	−0.7%
M0273 (Pic on/off)	−0.1%	0.0%	−0.2%	−0.0%	−1.7%	−0.3%
M0273 (PU on/off)	0.0%	−0.0%	0.1%	−0.0%	−2.1%	−0.4%

Consider for SNR only.

We would want multi-level switching – e.g. enable/disable at the SPS and SH level.

It was remarked that the text has problems (e.g. w.r.t. reference picture list construction).

It was remarked that a substantial portion of the average gain came from one sequence (People on Street), and generally only from one class of sequences (class A).

Some concern was expressed regarding the complexity of the technique.

It was commented that the complexity measurements reported do not seem entirely valid.

Track A recommended to adopt this into the SHM (for the SNR scalability case only, M0087 for TextureRL, M0273 for RefIdx using two reference indexes for the BL referencing, multi-level switch, subject to text review). Further discussion and possibly studying in a CE was then requested in plenary.

In later discussion, it was indicated that text had been provided and reviewed by some interested experts.

The scheme was characterized in further discussion as basically a denoising filter.

It was remarked that selective pre-processing of source material is another way to achieve denoising.

In further discussion, a variation that uses only one reference index for RefIdx was described, where HL syntax indicates whether the inter-layer referencing index refers to a filtered picture or not.

It was noted that the current RefIdx approach restricts MV values to 0 for inter-layer prediction (as an encoder restriction, not a syntax change). The presenter indicated that allowing non-zero motion vectors would not provide any significant benefit (with the current MC interpolation filter).

Almost half of the gain when using the two-index approach was noted to be from one sequence in the test set, and the gain was noted to be largest for low-delay P operation.

A participant remarked that having this might avoid more frequent use of bipred, which has higher complexity.

A participant remarked that the SNR scalability case is the most difficult variation in terms of complexity, since the BL has resolution equal to the EL.

It was noted that we might need not need BL filtering at all in the SNR case if we don't apply such a scheme. Otherwise, a "pointer-only" referencing method could apply.

A possible relationship to multi-view was noted. Multi-view referencing (as designed) does not use filtering when referencing other pictures of the same resolution.

A participant asserted that the gain may come from removal of quantization noise.

Suggestion of "Test model only" adoption (two reference index variant, without presumption of "automatic" promotion to WD). Disabled by default?

Further test in CE.