Organisation internationale de normalisation

5.4Core experiment #2: Adaptive reference filtering for MVC

5.4.1.1.1JVT-Z020 ( Prop 2.2/3.1) [P. L. Lai (USC), P. Pandit, P. Yin, C. Gomila (Thomson)] CE2: Adaptive reference filtering for MVC

This contribution reported work and results for CE2 (JVT-X302) to evaluate coding gain of adaptive reference filtering (ARF) for MVC. To avoid needing a two-pass encoding procedure (initial estimation, filter design, then encoding with filtered references), a fixed set of filters was reportedly designed by clustering the ARF filters with a method described in previous JVT documents (JVT-W065, JVT-X060, JVT-Y041). They were 3x3 filters with symmetric constraints as proposed in JVT-X060. The results provided in this document were for anchor-only sequence coding with P and B frames. The average bit rate savings for anchor-only sequence coding was asserted to be 5.69%, or equivalently 0.26 dB. The gains for anchor-only sequence coding were asserted to range from 0.07 dB to 0.45 dB. The asserted gain was larger for sequences with stronger focus mismatches such as Race1 and Rena.
Question: Are there applications that would use anchor-only coding? Response: Professional applications such as editing – note that there is a profile proposal for such a profile.
Question: What would the gain be when not using anchor-only coding? Seems very small.
Question: Is the gain additive to IC? Response: Not tested.
Question: Complexity impact? Extra filtering to be done during inter prediction.
Remark: Was combining the two filters considered.
Question: How is the reference filtering adaptive? A particular filter is assigned to each position in the reference picture list (no ability to change which filter is applied to which position). The only adaptivity is the dependence of the filtering on the reference picture list index.
JVT Disposition: Further study was encouraged.