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

4Core experiments in SHVC (24)

4.1SCE1: Arbitrary scalability ratio support (5)

4.1.1SCE1 summary and general discussion

JCTVC-O0031 SCE1: Summary report of SHVC Core Experiment on support for arbitrary scalability ratio [E. François, E. Alshina, J. Chen]

Discussed Wed 23rd afternoon (GJS & JRO).

SCE1 aims at evaluating different candidate upsampling filters for spatial scalability with any ratio between 1x and 2x. Two proposals are considered in this SCE.

Proposal	Initial proposal	Proposal documents	Cross-checking documents
1.1	JCTVC-N0219: Non-SCE1: On arbitrary spatial ratio scalability in SHVC	JCTVC-O0076 (Qualcomm, Samsung)	JCTVC-O0124 (Canon)
1.2	JCTVC-N0273: On the selection of fixed filters for upsampling	JCTVC-O0052 (Arris)	JCTVC-O0281 (Samsung)

The coding efficiency impact was minimal overall. A proponent noted that in some cases, JCTVC-O052 had a drop in performance. No benefit was shown for either of the alternatives to the current design (as documented in O0031 tables 2 and 3, first column).

For downsampling, both schemes seemed to work properly (non-normative). The current SHVC (informative) downsampling filter seems a bit better on average than the JSVM filter. Non-normative anyway.

This tested zero phase alignment for the upper left sample in the base and enhancement layers (as with an odd tap length filter).

Do we need to support arbitrary scalability ratios? Most participants said would not want to have multiple profiles on that basis.

Several participants said that the complexity impact for arbitrary scalability ratios is minimal (much less than with AVC) and it would be desirable to avoid the potential profile forking.

Proposal 1 – no significant difference compared to “reference” DCT-IF filters

Proposal 2 – small drop particularly for 1.5x scalability, whereas no benefit for arbitrary scalability.

Conclusion: If arbitrary scalability would be implemented, neither of the two proposals seems to provide a significant benefit.

Part of the CE was also about investigating the performance with different downsamplers

On average, the SHVC downsampler performs better than JSVM, whereas for some sequences this is not the case; generally, due to lower frequency cutoff, JSVM has lower base layer rate, and results may not fully be comparable. It can also be concluded from the results that both downsamplers are suitable for arbitrary scalability ratio.

Several experts expressed support for including arbitrary scalability ratio. Unlike in SVC, where this had implication on irregularity in deriving the partitions, it is a negligible burden in HSVC and might not justify definition of separate profiles.

One expert mentions that in case of hardware implementation potentially a burden might exist (as with low number of filters, multiplier-free implementation could be used).
Decision: Adopt ASR with filters as documented in O0031 tables 2 and 3, first column. Also update downsampler as tested for SHVC (see JCTVC-O0071) to enable ASR. (Update SHM document accordingly.)

(Reference position calculation unchanged.)

The burden of proof that ASR should not be supported in some profile would need to show strong justification of why it is difficult to support.
Note: There is a proposal to consider support of other phase shifts O0215.

Note: Chroma has a fixed alignment relative to luma in the draft (half shifted vertically).