International organisation for standardisation organisation internationale de normalisation
Yüklə 8,24 Mb.
|
5.3New Technical Contributions
m23558 (Disney) noted that the response to the CfP considered only multi-view video as input with synthesis performed at the receiver without supplemental information. This contribution proposed to shift data extraction and warp calculation to the encoder side. This requires the warping encoder to transmit the supplemental information. Warps are represented by a 2D quad-grid for each frame. One warp is generated for each view position. The proposed encoder applied a quincunx partitioning, followed by spatio-temporal prediction, quantization and entropy coding. It was reported that the coded warps require 3.7% on average of the total bit rate. The warp encode and decode times were reported, and compared with depth-based synthesis. It was recommended that such a warp encoder be considered for standardization within the AVC and HEVC compatible frameworks. The necessary software would include a warp encoder, decoder and view synthesis, but it was unclear whether warp generation software could be provided. One option for a CE would be to evaluate the efficiency of warp representation relative to disparity representation. Visible subjective gains would be needed to justify an alternative representation. It seems beneficial to study the differences/similarities of warp representation and disparity representation, and also to consider whether disparity maps could be used as input to the proposed warping process. Further study was encouraged. m23650 (Samsung) reported the coding efficiency of 2-view and 3-view input (without depth) within an HEVC-based framework, including all coding tools on versus block-level tools turned off (i.e. only using inter-view prediction with HL syntax). Reported gains were 28-38% with simple inter-view prediction relative to HEVC simulcast, and it was reported that block level coding tools only provide a 4% additional gain. It was suggested that the addition of block-level coding tools needs to be justified relative to performance. m23669 (Philips/Ghent) proposed using AVC for one view and HEVC for additional views and depth. Inter-view prediction is used based on decoded reference pictures. A gain of 31% was reported relative to 3D-ATM, and 34% loss was reported relative to using HEVC for coding all views. It was noted that no special coding tools had been implemented and that further gains are expected when such coding tools are introduced. It was suggested that a basic architecture with an AVC base layer and HEVC enhancement could be pursued on a faster standardization track (enabling only inter-view prediction, without using depth). Additional enhancements with block level coding tools and depth may also be considered. This was agreed – software integration was planned with 3D-HTM software. A workplan is needed considering other 3D-HTM plans. The timeline for standardization would be aligned with HEVC extensions. m23697 (Nokia) proposed a modificaiton of the slice header prediction in the current TMuC. Shortcomings of the current approach were summarized in the contribution, e.g. limited to decoding order where the depth component follows the respective texture component, inheritance issues, etc. The proposed "group of slices" (GOS) parameter sets could support any decoding order and enable more flexible prediction of syntax elements. There was support for further study on the proposed features. There was a question on maintaining compatibility with MVC that would need to be considered, as well as other design issues. m23721 (Nokia) considered reduced resolution depth map coding, in which the depth map image is currently up-sampled within the coding loop. Their study shows that no significant penalty on overall 3DV coding performance is observed when depth data is coded with integer-motion vector accuracy. The proposed to modify the DMVP to disable in-loop resampling of depths. The proposed mode of DMVP requires 75% fewer arithmetic operations than the original design and in-loop resampling is not required. Less than 1% loss in coding performance was reported. This was discussed in a BOG on low-complexity in-loop operations, and it was decided not to adopt this proposal at this time. However, further study was encouraged. m23723 (Nokia) proposed a flexible coding order in HEVC. Results were reported with 3D-HTM with certain sets of tools disabled. 11-14% loss was reported with flexible coding order enabled. The benefits of flexible coding order are not clear. Some of the tools may not be compatible with flexible order, and there also seem to be issues with bitstream extraction. There were concerns about the effort that is needed to support this. A BOG on Flexible Coding Order was established to hold discussions on this topic. A BOG report was produced (m23990). Recommendations:
Software integration plans were discussed. It was agreed to enable this feature in the software and make it available before the next meeting. m23731 (Zhejiang) proposed several modifications to mb_type coding for depth. Gains of 0.6% on average were reported for texture and depth. It was noted that we currently do not have any special modifications to entropy coding for depth data, and the gains reported by these simplifications did not seem to justify making a change to that module. m23807 (Qualcomm) contained three aspects:
m23872 (Sharp) proposed a modification in depth-based MV prediction that also uses average disparities of neighbouring blocks. This was claimed to be more robust to depth errors. Gains of 0.28% for C2 and 0.45% for C3 were reported with a few modifications for texture coding and no modification for depth coding. Conclusion:
Yüklə 8,24 Mb. Dostları ilə paylaş:
|