International organisation for standardisation organisation internationale de normalisation
Yüklə 4,35 Mb.
|
7.3New Technical Contributions
Software Integration & Interlaced m24731 (Sony): This contribution reported on the software integration of coding interlaced texture with progressive depth, which was decided to be supported in the working draft of the MVC extension at the previous meeting. The software is based on JMVC 8.5 and has been provided for verification purposes. Three possible structures were considered: (a) frame coded interlace for texture and depth frame, (b) first texture field with depth frame, then second field texture only, (c) first texture field only, then second texture field with depth frame in AU2. The third option is not integrated since there are issues with POC setting. The contribution recommended restricting case (c) in the working draft text and only allowing case (b). The corresponding text to support this proposed change was provided in the contribution. If it was desired, the proponents could plan to support this coding functionality in the ATM software; however it was noted that it would be necessary to add support for interlaced coding of texture. Decision: Adopt proposed text to PDAM. Discussion on software: The group discussed whether interlaced tool support should be added into ATM software at this time. This was considered undesirable by the group. It was agreed by the group that it is sufficient for this functionality to be provided through the independent JMVC based software implementation. m25069 (ITRI): This contribution presented a study of interlaced coding for texture and depth. It provides a survey of 3D camcorders on the market and noted that interlaced capture is supported in several products. Two experiments were conducted. Experiment 1 studies interlaced content coding in the JM context to enhance the coding efficiency for interlaced MVD (texture and depth). The contribution reports 13.5% coding gains when interlaced coding tools are used. Experiment 2 studies interlaced texture with progressive depth. The contribution reports 21.2% gains on synthesis results when using interlaced coding of both texture and depth. The proponents recommended that interlaced texture and depth be supported in the working draft. Discussion: It was remarked that there do not seem to be any simulation results with progressive coded depth at full frame rate. There was a question from the group on the source of interlaced depth. It was responded that interlaced depth may be estimated directly from texture fields. The group encouraged further study be done with interlaced MVD sequences. The experts suggested collecting interlaced stereo test material, including camera parameter information. Inclusion of interlaced depth into the specification should also be studied. m24916 (Nokia): This was an informative contribution that reports on support for flexible coding order in the HTM software to enable further testing and evaluation of tools that require alternative coding dependencies. It was reported that there is no coding penalty on texture coding, but there is a 1.65% coding loss on the total coded data, and a 1.4% loss on the synthesized result. The group agreed that this is a useful configuration to support in the software during the development phase. m25013 (Poznan): It was reported in this contribution that all of the Poznan-proposed tools have been integrated into 0.3 and 0.4 versions of the HTM software. The list of tools and their evaluation status is summarized below:
All of the coding tools were tested by HHI and they reported results with a perfect match; there were no problems with source code, implementation or simulation. The tools are currently available for further evaluation and all tools are configurable. It was verbally noted by a member of the group that when all the tools are switched on, there is a coding loss. It was suggested that some offline discussion is needed to determine the CE plans with respect to further evaluation of these newly integrated tools. This discussion was later revisited. It was considered by members closely involved in this activity that all technologies under consideration are planned to be investigated further as part of the new CEs to be established at this meeting. The following modifications to the CTC were agreed:
It was suggested by a member of the group that a document on detailed software integration plan and guidelines be considered. A first draft was provided later in the week by Gerhard and Dmytro. This document was reviewed by the group, and the group agreed to have this document as an output document of the meeting. Normative coding tools m24766 (LG): The main idea presented in this contribution is that the temporal predicted residual is not used to predict the inter-view predicted residual. Simulations were performed in which the proposed method is only applied to uni-directional prediction and both uni- and bi-directional predictions. The contribution reports a gain of 0.1% on average for both cases with 0.3-0.4% gains on two dependent views. The method has been cross-checked by HHI in m25030. The group discussed whether to adopt this proposal, study it further in a CE, or encourage further study outside of a CE. Decision: Adopt m24766 m24938 (Qualcomm): This contribution presents a proposal on inter-view residual prediction that avoids pixel-level operations at the parsing stage. The availability of inter-view reference pictures (IVRP) are marked based on the CBF and coding modes of a residual block. It was reported that no coding loss was observed. The cross-check report by Aachen in m24963 confirms this result. The overall encoding and decoding time is not significantly changed, but it is claimed that the hardware implementation is significantly better. As a general point, it was also noted by a member of the group that parsing should be independent of any signal processing on decoded samples – it is not clear that the current IVRP tool fully satisfies this design principle. There was also a question on whether dependence on syntax elements in other slices (views) is desirable, but this may be reasonable since if the base view is corrupt the non-base views may have limited use anyway. Further study seems desirable to clean-up the design. Decision: Adopt m24938. m24816 (Samsung): This contribution presented a proposal on inter-view coding with adaptive luminance compensation. Several non-normative and normative modifications were made in the ATM software. For instance, only synthesized frames were used as a source of prediction with limited search range, and only 4x4 transform was enabled. An additional flag is added at the MB level to signal that this method is used. The contribution reports 0.75% gain for coded data, and 0.39% gain for synthesized result. Results have not been cross-checked. It was noted that the gains are higher for non-base views. It was reported that the encoder complexity increases by approx. 20%, while decoder is reduced by 28% (the proponents noted that they need to recheck the decoder result). It was noted that there is no additional side information transmitted. The proponents recommended establishing a CE on luminance compensation or to evaluate this method within CE1 on view synthesis prediction. It was noted by the proponents that the method can be extended to inter-view prediction as well, but this has not yet been tested. It was mentioned by a member of the group that this proposal is related to the previously proposed illumination compensation tool, which explicitly sends illumination differences. The group agreed to study this tool further in CE1, with a slightly broader scope to include evaluation of ALC for inter-view prediction as well. m25005 (LG): This contribution proposes a modified truncated Golomb-Rice code for coeff_abs_level_minus3 coding in the current CABAC design of HEVC. The proposal modifies the truncated Golomb-Rice code based on the statistical characteristics of the quantized transform coefficient. It is reported that the proposed method provides approximately 0.1% BD-rate savings for two dependent texture views in the 3-view case without depth coding. It was noted by members of the group that this is an active topic of study within JCT-VC and it is not likely that the 3DV group would open up study on this topic. Also, it is unclear that 3DV should alter the design of the entropy coding module. m25008 (LG): This proposal describes a simple warp map converter for warp map coding. The contribution suggests a compatible coding approach for warp data proposed by Disney. Coding comparisons with the Disney approach were provided. In the experiments, the resolution of the warp map was 180x100 and the overall bit rate was very low. The proponents recommended integrating the converter/reconverter into the existing software frameworks. This contribution was considered informative and the group encouraged further study on this topic. m25050 (LG): This contribution explores the compatibility between image domain warping (IDW) proposed by Disney using texture and depth coded data. The contribution proposed to use depth maps to facilitate IDW view synthesis methods within the current 3DV framework. This contribution was considered informative and the group encouraged further study on this topic. Non-Normative m24765 (Peking/LG): This contribution aims to detect zero distortion change in a synthesized view before the rendering process. The proposal reports an encoding time reduction of 10-20% with no loss in BD performance. A cross-check of results was provided by Samsung in m25059. It was concluded by the group to study this method along with other VSO simplifications as part of CE8. m24790 (KWU): View coding order is specified as part of the CTC and only forward uni-directional prediction is considered for inter-view prediction, i.e., inter-view reference pictures are only available in reference picture list 0. This proposal asserts that further gains are possible when inter-view references are also added to list 1. The current software can be configured to achieve this with a change in the GOP string. The contribution reports a 13% coding gain on non-base views for texture coding when the method is tested in HTM 0.4. A cross-check of the results was provided by Samsung in m24881. The proposal recommends that the CTC be changed in order to make inter-view references available in list 1. Discussion: It was pointed out by a member of the group that the latest version of HM supports a combined list, and as a result, the total gains of the proposed method may be less. It was suggested by the group that this configuration also be tested in context of the ATM software. It was concluded by the group that no change should be made to the CTC at this time, but further study was encouraged. m25003 (LG): This contribution proposed to reduce the maximum RQT for depth coding. The proposed algorithm shows a 0.1% loss in coding with 12% reduction in encoding time. Discussion: It was noted by a member of the group that the current specification and software already supports encoding in this manner. This appears to be an encoding configuration setting and not necessarily a subject for standardization and does not require further study. If there are normative aspects for RQT depth, the group may further consider the subject. m24807 (EPFL): This contribution presented a study of 3D quality with different metrics: PSNR, SSIM, MS-SSIM (Multi-Scale Structural Similarity Index), VSNR (Visual Signal to Noise Ratio), VIFp (Visual Information Fidelity, pixel domain version), WSNR (Weighted Signal to Noise Ratio), PSNR-HVS (PSNR Human Visual System), PSNR-HVS-M (PSNR Human Visual System Masking), DCTune, VQM (Video Quality Model). Some observations from the contribution include:
Quality assessment scores from the 3DV subjective tests were used to determine the best metrics. From the results, it was suggested to consider VIFp, VQM, MS-SSIM, or SSIM as part of the evaluation in Core Experiments, as they give much better results than PSNR, if evaluation is performed on synthesized views. The group did not agree to any changes in the evaluation process at this time, but encouraged CE coordinators to consider alternative metrics as part of their workplan. In general, further study on the topic of objective quality assessment metrics is needed. m24889 (Nokia): This contribution presented a number of suggested changes as an editor’s input to the AVC-based working draft. m25146 (IT): This was an informative contribution on coding for 3D holoscopic video. A primary objective of this system is to provide continuous parallax and eliminate "cardboard cut-out" effects in the rendered 3D video. The document describes the recording of scene objects with special cameras using normal lenses, as well as with a micro-lens array for 3D holoscopic visualization. The conversion of content to multi-view display was presented, where the holoscopic data gives both vertical and horizontal parallax. HEVC-based intra-coding was performed for the holoscopic images, where strong cross-correspondences in the holoscopic images are exploited. Experiments are conducted with HM4.0 and HM4.0 with a self-similarity estimation method. A scalable coder was developed with a 2D base view, a first enhancement layer for second stereo view, and a second and third scalability layer for full holoscopic image content. Yüklə 4,35 Mb. Dostları ilə paylaş:
|