JVET-G0026 Polyphase subsampling applied to 360-degree video sequences in the context of the Joint Call for Evidence on Video Compression [A. Gabriel, E. Thomas (TNO)]

Discussed Thursday 16:45 (GJS & JRO).

Bitstreams were not submitted for this contribution. The proponent said that the results were varying substantially from sequence to sequence, so there is some lack of maturity in the technology at this point.

This contribution presents an analysis of the compression efficiency of 360° video sequences transformed via the polyphase subsampling technique presented in JVET-B0043. This may be conceptually similar to the H.261 Annex D graphics sampling scheme.

This proposed technique decomposes the input video signal (luma and chroma components) by subsampling into lower resolution descriptions using a polyphase subsampling. The multiple lower resolution versions of the signal are then encoded in the same video bitstream via a temporal multiplexing. This enables the decoder to select and to decode the appropriate number of resolution components for the desired output resolution. The polyphase subsampling technique without low-pass filtering is known to be potentially create artefacts such as aliasing when high spatial frequency is present in the signal. The first reported experiment is the comparison of 360 4K video sequences encoded with HM against the polyphase subsampled sequences encoded with HM. The results reportedly show that the best Y-PSNR BD-rate impact is −67.7% for the sequence SkateboardingInLot while the worst is 86.9% for the sequence Trolley. The average Y-PSNR BD-rate impact over all the sequences is 14.3%. The second experiment reported is a simulcast comparison, that is the comparison of 360 4K video sequences encoded with HM plus the 360 video sequences downscaled in HD encoded with HM the against the polyphase subsampled sequences encoded with HM. The results reportedly show that the best Y-PSNR BD-rate impact is −76.50% for the sequence SkateboardingInLot while the worst is 31.30% for the sequence Trolley. The average Y-PSNR BD-rate impact over all the sequences is reportedly −19.10%.

It was remarked that some of the inconsistency in the results may be from bit rate matching effects.

A participant highlighted that the measured performance is reported from measurements made directly in the ERP domain, not considering the spherical domain.

It was commented that since the I frame refresh is not altered to compensate for the 4 coded pictures used per high-resolution frame, the I frame frequency is effectively four times higher than for the anchors, which could adversely effect the results. The QP variation would also affect different spatial positions differently.

The contributor remarked that the encoding time was faster with the scheme than without it. A participant commented that this may be due to the motion search early termination. Each MV value is effectively doubled.

It was commented that the aliasing in the sampling structure might have subjective quality issues.