International organisation for standardisation
Adaptive loop filter (“Wiener”)
Yüklə 9,08 Mb.
|
6.5.2Adaptive loop filter (“Wiener”)6.5.2.1.1.1.1.1.1JCTVC-F076 ALF coefficient prediction [K. Andersson (Ericsson)] Prediction of filter coefficients – adds some complexity – very small gain. Used for the chroma in current HM. (There is only one filter in the chroma case). Predicts the value of center tap based on the value of other taps, and codes the difference. For luma, the filter tap values are predicted from other filter values. Might be helpful to keep in mind in the future, but let's not tinker with it in this way at the moment. Other aspects should be stabilized before worrying about this. 6.5.2.1.1.1.1.1.2JCTVC-F161 Complexity Scalable ALF [M. Li, P. Wu, Z. Li, W. Zhang (ZTE)] Proposes to have varying complexities of ALF that can be enabled (e.g., for different profiles or for particular picture patterns). Benefit was shown relative to LC (where the anchor does not use ALF). A participant remarked that the PPS is not the right place to put syntax with a larger scope. The idea of a "slice parameter set" was also noted. It was remarked that the current syntax is flexible enough already to support some forms of this. The motivation was described as partly a matter of "decoder complexity scalability". Contribution noted. 6.5.2.1.1.1.1.1.3JCTVC-F209 Cross-check of complexity scalable ALF (JCTVC-F161) [J. Xu (Microsoft)] [late upload 07-12]
Performs Laplacian edge detection to control filter selection. The proposal suggests changing the number of categories for the selection. Gain is small, but it was suggested that further study might uncover a benefit from optimizing the selections. 6.5.2.1.1.1.1.1.5JCTVC-F271 Grid displacements for in-loop filtering [S. Esenlik, M. Narroschke, T. Wedi (Panasonic)] This contribution proposes a method to reduce the complexity of the LCU- (Largest Coding Unit) based processing of Sample Adaptive Offset (SAO) and Adaptive Loop Filter (ALF) operations via displacements in the filter grids. With the contribution the “Filter Design Window” in the encoder and “Filter Application Window” in decoder are matched in order to achieve uniformity in filtering within a “Filter Application Window”. Moreover the new placements of the ALF and SAO grids reduce the line memory requirement for in-loop filtering by 3 lines, which corresponds to 25% reduction. It was noted that this applies only to the "region based ALF" scheme, and may not help the worst case since the worst case may not use that ALF scheme – it could use CU based ALF The goal is to save decoder complexity. The proposal was somewhat modified between versions of the proposal during the meeting. Further study in a CE was suggested. 6.5.2.1.1.1.1.1.6JCTVC-F683 Cross-verification of Panasonic's proposal JCTVC-F271 [M. Ikeda, T. Suzuki (Sony)] [late reg. 07-07, upload 07-08] Software was analyzed as well as operated. 6.5.2.1.1.1.1.1.7JCTVC-F320 Subjective tests on ALF and SAO [O. G. Sezer, M. Budagavi (TI)] This contribution presents results of informal subjective tests conducted on two loop filtering operations in HM-3.0: SAO and ALF. Test videos were presented at their actual frame rates to the viewers. Four configurations were tested using low delay B high efficiency (LB-HE) condition: HM-3.0 Anchor v/s SAO-off+ALF-off, HM-3.0 Anchor v/s SAO-off, HM-3.0 Anchor v/s ALF-off, HM-3.0 Anchor v/s ALF-Luma-off. Overall, the subjective difference between the HM-3.0 Anchor and the test configurations is less than expected. Among these four configurations, ALF-Luma-off gave subjective results closest to the HM-3.0 Anchor. This contribution requests that JCTVC conduct subjective testing of ALF and SAO in a core experiment setting to evaluate the subjective gains provided by the tools and decide on the minimal set of configurations that best provide subjective quality improvements so that the minimal set can be included in the final standard. The report is not from a large rigorous study. The impression is that SAO provides a larger visual gain than ALF; when SAO is enabled, ALF is not providing very substantial further improvement. Some other participants indicated that they had done some similar experiments and not reached the same conclusion. The BoG coordinated by M. Zhou (JCTVC-F763) was asked to try to do some informal investigation of this during the meeting. Further study was encouraged. 6.5.2.1.1.1.1.1.8JCTVC-F342 ALF Complexity Analysis [T. Hellman (Broadcom)] This submission examines the implementation complexity of the adaptive loop filter (ALF). It reports that the worst-case ALF complexity, as presently specified, is on the order of 130 operations (multiplies and adds) per luminance pixel, or roughly 16 billion operations per second for an HD video stream. It compares this complexity to that of the inverse transform and motion compensator, and recommends study of ways to reduce this complexity. The proponent indicated that the changes being adopted at this meeting is making substantial progress at reducing the burden. 6.5.2.1.1.1.1.1.9JCTVC-F498 Adaptive Loop Filter Merge in Temporal Domain [X. Zhang, R. Xiong, S. Ma, W. Gao (Peking Univ.)] This contribution proposes a temporal adaptive loop filter (ALF) merging method. The proposed method utilizes decoded ALF parameters from prior frames to implement the adaptive loop filtering. The usage of whether prior decoded filter bank or the current filter bank in HM3.0 is signaled in the bitstream. Compared with HM3.0, the proposed method reportedly achieves 0.3%, 0.2%, and 0.2% bit rate reductions for HE-RA, HE-LD, and HE-LD (P), respectively. The encoding time is reportedly increased by 2%, while the decoding time is reportedly increased by 2~5%. Coding results are confirmed by Microsoft with JCTVC-F207. Similar to JCTVC-F321, but a bit less flexible regarding which picture to obtain prior coefficients from. 6.5.2.1.1.1.1.1.10JCTVC-F207 Cross-check of adaptive loop filter merge in temporal domain (JCTVC-F498) [J. Xu (Microsoft)] [late upload 07-13] 6.5.2.1.1.1.1.1.11JCTVC-F542 ALF with low latency and reduced complexity for HEVC [A. Fuldseth, G. Bjøntegaard (Cisco)] The document describes an ALF for low latency applications. Preliminary simulation results are presented. ALF as specified in HM 3.0 brings a coding gain of typically 4-6% measured in BDR. The operations are mainly performed on a frame basis. This may lead to additional latency caused by the coding process. This makes ALF less attractive for low (sub-frame) delay real time applications like video conferencing. Furthermore ALF represent a considerable implementation complexity. The present document therefore describes some ideas and initial simulation tests on an ALF version that could be useful in connection with low latency applications. HM 3.0 perform ALF on a frame basis. According to the present document ALF is performed on smaller units. 64x64 and 128x128 luma pixels have been used. This will be referred to as an ALF-unit. The ALF coefficients are quantized more coarsely than in HM 3.0. This is mainly in order to save coding bits since coefficients may have to be sent for every ALF-unit. For the 5x5 diamond shape structure 7 ALF coefficients are calculated. With the coarse quantization, especially the DC response of the filter may suffer. To overcome this only 6 out of 7 coefficients are quantized and transmitted. The last coefficient – representing the center position of the filter is calculated to give a final DC response of 1. Calculation and quantization of a set of filter coefficient is performed for an ALF-unit. Then an RD test is made to decide if the filter shall be used. 1 bit is used to signal whether the ALF is enabled or disabled. If the outcome of the RD test is positive, filtering is performed with dequantized filter coefficients. No further switch between filters is performed.
The proponent was asked about subjective quality, and indicated that subjective quality is what led to the interest in the subject. The scheme, as proposed, is substantially less complex and has much more compact source code than the current ALF design. Further study in a CE was suggested. 6.5.2.1.1.1.1.1.12JCTVC-F522 Enhancing block/region based Adaptive Loop Filter by MediaTek, Qualcomm, Sharp and Toshiba [I. S. Chong, M. Karczewicz (Qualcomm), T. Yamakage, T. Watanabe, T. Chujoh (Toshiba), C.-Y. Chen, C.-M. Fu, C.-Y. Tsai, Y.-W. Huang, S. Lei (MediaTek), T. Ikai, A. Segall, T. Yamamoto, Y. Yasugi, T. Yamazaki (Sharp)] This contribution presents a merged version of two CE8 subtest 1 proposals (JCTVC-F321 and JCTVC-F384), which aims to improve block/region adaptive loop filter (ALF). JCTVC-F321 proposes temporal coefficient prediction and BA enhancement, and JCTVC-F384 proposes flexible class representation and additional block-based adaptive mode. It is reported both proposals improve coding gain, and the gain of both proposals are additive. Coding gains of 0.3%, 0.7%, 0.6% and 0.7% in HE-AI, RA, LB, and LP were reported. The decoding time increased by 3% on average with no changes in encoding time. See discussion above in sections discussing JCTVC-F321 and JCTVC-F384. 6.5.2.1.1.1.1.1.13JCTVC-F182 Cross-check on ALF proposal JCTVC-F522 [K. Sugimoto, S. Sekiguchi (Mitsubishi)] [upload 07-16] Coding results were confirmed by Mitsubishi. 6.5.2.1.1.1.1.1.14JCTVC-F465 Item 8 (Region-based ALF only) of Experiments on tools in Working Daft (WD) and HEVC Test Mode (HM-3.0) [I.-K. Kim, E. Alshina, J. Chen, T. Lee, W.-J. Han, J. H. Park (Samsung), V. Sze, M. Budagavi, M. Zhou (TI)] Item 8: Region-based ALF only. Just for information – for further study – no action, although it would be interesting to reduce the number of modes. 6.5.2.1.1.1.1.1.15JCTVC-F465 Item 9 (Single pass ALF) of Experiments on tools in Working Daft (WD) and HEVC Test Mode (HM-3.0) [I.-K. Kim, E. Alshina, J. Chen, T. Lee, W.-J. Han, J. H. Park (Samsung), V. Sze, M. Budagavi, M. Zhou (TI)] Item 9: Single pass ALF. Information on 1 and 2 pass behaviour within the context of HM4 is requested (do in CE8). 6.5.2.1.1.1.1.1.16JCTVC-F465 Item 10 (One filter only for ALF) of Experiments on tools in Working Daft (WD) and HEVC Test Mode (HM-3.0) [I.-K. Kim, E. Alshina, J. Chen, T. Lee, W.-J. Han, J. H. Park (Samsung), V. Sze, M. Budagavi, M. Zhou (TI)] Item 10: One filter only for ALF. For further study.
Yüklə 9,08 Mb. Dostları ilə paylaş:
| ||||||||||||||||||||||||||||||