International organisation for standardisation

6.8.4Reference picture list management

6.8.4.1.1.1.1.1.1JCTVC-F549 Coding with a single, unified reference picture list [D. Flynn, N. Šprljan, M. Mrak (BBC)]

The HM software has been modified to support coding with a single unified reference picture list. The concept of reference pairs is introduced, where a pair consists of either two references (bi-predictive coding) or one reference and a null element (uni-predictive coding). The pairs are ordered and a single codebook is used for coding, which avoids separate L0 and L1 indexing. The main benefit of this solution is that the selection of reference frames in bi-prediction is not restricted to all combinations of reference frames from two lists (as per WD3). Instead, it is suggested that we could now use a smaller, equal or larger set of reference frame combinations. This solution reportedly reduces the number of lines in both the WD and HM, while providing a marginal RD performance increase of ~0.1% when the software is configured to emulate the equivalent of the HM-3.0 reference pairs.

Default list construction and reference list modification schemes are needed for this.

The number of possible pairs is large when there is a large number of pictures in the buffer (on the order of the square of the number of pictures in the buffer). For example, in AVC there can be N=16 pictures in each list, which would produce N+(N^2)/2 = 144 combinations for uni-prediction and bi-prediction.

Combining this with explicit weighted prediction might need some thought.

The proponent indicated that the proposal is not necessarily fully mature at this stage (although there is no coding loss with the common conditions).

A cross-checker said that the implementation is straightforward and simplifies the software.

It was suggested that the parsing method may cause entropy decoding difficulties and loss robustness issues.

Establish an AHG to think further about this.
6.8.4.1.1.1.1.1.2JCTVC-F645 Cross-check report on Coding with a single, unified reference picture list (JCTVC-F549) [I.-K Kim (Samsung)] [late reg. 07-05, upload 07-06]
6.8.4.1.1.1.1.1.3JCTVC-F655 Cross-check report on unified reference picture list (JCTVC-F549) [J. Jung, J. Le Tanou (Orange Labs)] [late reg. 07-05, upload 07-11]
6.8.4.1.1.1.1.1.4JCTVC-F573 The construction of combined list for HEVC [S. Fang, Y. Yu, L. Wang (Motorola Mobility)]

This contribution proposes a default list construction order for the combined list for uni-prediction (as in the current design). It is reportedly more efficient if there are multiple consecutive non-reference pictures, without increasing complexity.

The first element is to suggest that the default uni-prediction list order is to order by increasing POC distance. If multiple pictures have equal temporal distance, average QP is used as a tie-breaker.

Remark: The second aspect requires specification of a QP averaging computation.

Remark: If you lose a slice, the average QP becomes wrong, and list construction for subsequent pictures becomes impossible.

Remark: Using POC also has a loss resilience issue, as the loss of a picture may result in the loss of its POC.

Remark: Reference picture list modification can be used to provide robust ordering.

Measured benefit is very small.

Can consider this idea (and refinements of it) in the AHG that studies reference picture list construction.
6.8.4.1.1.1.1.1.5JCTVC-F627 Cross-check of Motorola Mobility’s proposal on combined reference list construction (JCTVC-F573) [C. Yeo (I2R)] [late reg. 07-04, upload 07-04]