International organisation for standardisation organisation internationale de normalisation
Yüklə 8,63 Mb.
|
- Bu səhifədəki naviqasiya:
- CE contributions
- Related contributions
CE4: Residual prediction
4.2.1.1.1.1.1.1.81JCT3V-C0024 3D-CE4: Summary Report on Residual Prediction [L. Zhang] One mandate of this CE was to investigate the removal of parsing issues in IVRP. It was reported that a 0.5% coding loss on video only would result. Two CE4 proposals have been submitted: JCT3V-C0049 and JCT3V-C0138. Comparison versus CTC anchors
There are also two CE4 related proposals: JCT3V-C0098 and JCT3V-C0139.
4.2.1.1.1.1.1.1.82JCT3V-C0049 3D-CE4: Advanced residual prediction for multiview coding [L. Zhang, Y. Chen, X. Li, M. Karczewicz (Qualcomm)] Inter-view residual prediction is enabled in the current HTM design to code the residue of dependent texture views more efficiently. This proposal is a follow-up of JCT3V-B0051 with some complexity reduction parts introduced. In this proposal, a residual prediction (ARP) is proposed to further improve the coding efficiency of inter-view residual prediction. In ARP, to ensure high correlation between residues of two views, motion of the current block of picture in current view is applied to the corresponding block in a reference view picture to generate residual in the base view to be used for inter-view residual prediction. Moreover, an adaptive weighting factor is applied to the residue signal so that the prediction error is further reduced. It is reportedly shown that the proposed method achieves compression efficiency gain of 0.9% and 1.4% for coded views in terms of BD bit rate compared to HTM anchor and CE4 anchor, respectively. Parsing dependency is removed by signalling IVRP unconditionally a CU level (not conditional at PU level), provided the reference picture is an inter-view picture (which is dependent on reconstruction of merge candidate list). (0.1% loss) Motion vector of current block is applied to reference block to generate the residual Additional motion compensation is needed – more complex (also increased memory bandwidth) – comes from different access unit. Weighting can be 0, 0.5, 1. Bilinear filter is used in the disparity compensation of the residual. CE4 anchor is modification of CTC without parsing dependency (parsing dep. is in effect in CTC) Q: Can it happen that the temporal reference picture of current view is not used in the co-located reference view? Yes, but then ARP would be disabled. Memory bandwidth is of concern and requires more thorough analysis (this also applies to other parts of HTM). Even though several experts expressed that the coding gain of this proposal would be desirable, concerns were raised with regard to additional complexity and the tradeoff versus the gain. No consensus was reached that this should be adopted; further study with more detailed analysis of complexity impact and possible further reduction would be desirable.
This contribution presents CE4.h results on the inter-view residual prediction (IVRP) in JCT3V-B0093, where two methods are proposed. One method is to solve the parsing dependency problem by removing the conditionally explicit signalling of the IVRP enable flag and applying IVRP to only CUs that select the temporal inter-view merging candidate of merge or skip mode with 2Nx2N partition. The other method is to reduce the complexity of IVRP by removing residual interpolation. The experimental results reportedly show that the proposed methods do not cause any coding efficiency loss. Removing parsing dependency by using residual prediction unconditionally for 2Nx2N with merge/skip gives very small gain (0.1/0.2 on the dependent view, 0+ overall). Removing interpolation of residual gives small loss (0.1% for coded view). Combining both gives zero loss. Decision: Adopt the part that removes the parsing dependency. No action on removing interpolation of residual.
4.2.1.1.1.1.1.1.86JCT3V-C0098 3D-CE4.h related results on simplification of residual prediction [J. Y. Lee, H.-C. Wey, D.-S. Park (Samsung), J. An, K. Zhang, J.-L. Lin, S. Lei (MediaTek)] This contribution proposes an asserted simplification of inter-view residual prediction. The residual prediction is only performed in dependent view when the corresponding block in base view is temporally inter-coded. Since temporal and inter-view predicted residual have low correlation, the temporally predicted residual of the current block generated by the temporal prediction is only re-predicted through the inter-view residual prediction. When both temporal and inter-view predictions are performed in bi-prediction case of B slice, the residual prediction is half enabled. However, the proposed method disables the residual prediction to avoid some multiplication operation, when the prediction types are different in the bi-prediction case. In addition, the proposed method removes an interpolation process in the residual prediction to reduce the complexity. The proposed method reportedly simplifies the residual prediction without the coding loss. No presentation necessary per removal of parsing dependency by adopting JCT3V-C0138. 4.2.1.1.1.1.1.1.87JCT3V-C0185 3D-CE4.h related: Crosscheck results on simplification of residual prediction (JCT3V-C0098) [S. Shimizu, S. Sugimoto (NTT)] [late] 4.2.1.1.1.1.1.1.88JCT3V-C0139 3D-CE4.h related: Mode dependent advanced inter-view residual prediction [J. An, K. Zhang, J.-L. Lin, S. Lei (MediaTek)] This contribution proposes a method to take advantage of both concept in JCT3V-B0093 and JCT3V-B0051 to enable the Inter-View Residual Prediction (IVRP) only when the temporal inter-view merging candidate in merge or skip mode is selected with 2Nx2N partition and the advanced residual generation is used. With this proposed method, the parsing problems are totally removed, the encoder RDO process is simplified with 7% encoding time reduction and significant BD-BR gain around 3% for v1 and v2 is achieved. No presentation necessary per removal of parsing dependency by adopting JCT3V-C0138.
| ||||||||||||||||||||||||||||||||||||||||||||||||||