JVET-O0549 AHG10: Encoder-only GOP-based temporal filter [P. Wennersten, R. Sjöberg, J. Enhorn (Ericsson)]

This contribution was discussed along with those in section 6.15 on Friday 5 July 1800–2130 in Track B (chaired by JRO).

This contribution proposes an encoder-only temporal filter. The filtering is done at the encoder side as a pre-processing step. Source pictures before and after the original picture to encode are read and an 8x8 block-based motion compensation method relative to the original picture is applied on those source pictures. Samples in the original picture are temporally filtered using source picture sample values after motion compensation. Only pictures at temporal sub layers 0 and 1 are filtered.

A similar method was proposed as JCTVC-AI0023 and adopted to HM at the previous JCT-VC meeting.

The method was reportedly tested under VTM-5.0 RA, LDB and LDP common test configurations. The average Y/U/V BD-rates for the common test conditions (CTC) are reported to be −3.49%/−6.96%/−6.53% (RA), −1.00%/−1.24%/−2.24% (LDB) and −1.47%/−1.95%/−2.78% (LDP). All BDR numbers were computed using unfiltered source sequences. The method is not proposed for AI encoding.

Version 2 of this document contains updated BDR numbers for RA.

It is proposed to adopt the proposed method into the VTM software but to disable it in the CTC.

Filtering strength is QP adaptive.

The proposal received support by several experts expressing interest.

This should not be in CTC (as preprocessing is usually undesirable for comparison), but interesting for experimentation. Could also be used for...