International organisation for standardisation organisation internationale de normalisation
Yüklə 8,24 Mb.
|
5.6.3Adaptive resolution coding5.6.3.1.1.1.1.1.1JCTVC-H0234 AHG18: Design considerations for Adaptive Resolution Coding (ARC) [T. Davies (Cisco)] This contribution considers various ARC features from a complexity perspective. It proposes level constraints on the number of reference resampling operations per frame period and the frequency with which resolution can be changed. A variety of short-length reference resampling filters were evaluated on low delay configurations, and their effects were reported to be transient, especially for up-sampling. For reference up-sampling, both simple linear and DCTIF upsampling filters reportedly had no loss on average against an anchor employing common filtering for reference and input/output pictures. For reference down-sampling, a 3-lobe Lanczos-windowed filter was observed to perform best of those tested against the anchor with mean luma BD bit rate losses of 0.45%. Particularly for high QP, coding gains were reported previously in some cases in JCTVC-G264. Constraints were proposed for how often the resolution could change (in output order) and how many reference pictures could be resampled when it happens. It was commented that it may be beneficial to constrain the resolution changes in output order as well as in decoding order. It is proposed for collocated motion vectors to be marked as not available when resolution differs. It was remarked that unless the higher resolution uses a coded picture size that is a multiple of two times the SCU size, there would be different alignment effects at the border. In the proposal software, padding was proposed. Various short filters were tested for the normative reference resampling filters and the non-normative encoder-only filtering. Only 2:1 resolution changes were tested. It was reported that for the upsampling filters, it basically didn't matter what the filters were, and a co-sited linear interpolation filter was suggested. For downsampling, it mattered a bit more, and a 3-lobe Lanczos filter (5 non-zero taps [-1 0 9 16 9 0 -1]) was suggested. A participant suggested alternatively considering the resampling filters that had been proposed in JCTVC-G264. It was commented that such a feature, if adopted, might need to be left out of some profiles (e.g. not used for broadcast). It was commented that the relationship with SVC should be studied. The proponent asserted that ARC is a simpler approach than SVC, and may have better coding efficiency. 5.6.3.1.1.1.1.1.2JCTVC-H0185 Cross-check of Adaptive Resolution Coding (ARC) [G. Van Wallendael, S. Van Leuven, J. De Cock, R. Van de Walle (Ghent Univ.)] [late]
This contribution presents a method for implementing Adaptive Resolution Coding (ARC) in which the DPB holds one coded size for each picture, which is its decoded size. The resampling and motion compensation filtering are combined in the decoding process (which lengthens the filter). It is reported that the scheme has encoding gains, relative to a different approach for dynamic resolution change, of between 0.1% to 1.8% in Luma and up to 4.9% in Chroma on average depending on configuration cases and scenarios (low-to-high or high-to-low). The DPB memory size is reportedly reduced at the decoder side, compared to previous ARC proposals, by avoiding storage of more than one resolution for each picture in the DPB. At the encoder side, both resolutions are still necessary to compute motion estimation only. The proponent indicated an interest in investigating the use of shorter filters. Further study would be needed to determine the potential value for such an approach. 5.6.3.1.1.1.1.1.4JCTVC-H0351 AHG18: Cross-check of JCTVC-H0321 by Cisco [T. Davies (Cisco)] [late]
(Informative contribution.) This contribution proposed the consideration of reduced resolution update (RRU) techniques (e.g. as in H.263 Annex Q) as potential methods for enhancing encoding performance for video compression. RRU techniques allow encoding a video signal by, commonly predicting that signal at full resolution while encoding any residual information at a lower resolution. Such information needs to be up-sampled before it is used for the reconstruction of the final image. It has reportedly been shown in the past that RRU methods can provide considerable coding advantages, both subjectively and objectively, especially at lower rates, versus full resolution as well as reduced resolution encoding. The proponent indicated that, unlike what was reported in previous contributions on the same topic, additional coding advantages or flexibility may be achieved by considering applying RRU concepts separately at different colour components or colour component groups, as well as extending RRU methods at the slice or region, and even at the coding unit level. Support also of down-sampling of not only the residual but also of the prediction signal was suggested to also be considered. This was asserted to essentially make the RRU process a superset that already potentially accounts for the techniques already presented in AHG 18. It was suggested by the submitter that such methods should be carefully evaluated. It was noted that there is some overlap in concept between RRU and the use of large block sizes (for MC and residual) and possibly also with selective dropping of high-frequency coefficients. Yüklə 8,24 Mb. Dostları ilə paylaş:
|