6Explorations – High-Performance Video Coding
A large quantity of video material is already distributed in digital over broadcast channels, digital networks and packaged media. More and more of this material will be distributed with increased resolution and quality demand. Technology evolution will soon make possible the capture and display of video material with a quantum leap in quality (temporal and spatial resolution, color fidelity, amplitude resolution). Networks are already finding it difficult to carry HDTV resolution and data rates economically to the end user. Therefore, further data rate increase will put additional pressure on the networks. Therefore a new generation of video compression technology that has sufficiently higher compression capability than the existing AVC standard in its best configuration (the High Profile), is needed. A study has been started on the feasibility of HVC, which is mainly intended for high quality applications, and a Call for Evidence is under preparation.
Before the Maui meeting, AVC anchors that shall be compared against prospective new technologies were provided as announced in the Draft Call for Evidence (see m16462, m16463). First investigations on the anchors were made (see m16375, m16463) and continued during the week. As some of the AVC anchors already seemed to be at subjectively transparent quality, it was decided to adapt the QP values for some of the test cases. Further, as one input contribution reported that the search ranges of motion estimation might be insufficient for some of the cases, the corresponding parameters were also changed. One contribution for an additional set of test sequences (1920x1080, 24fps, 10 bit RGB) was received (m16472). Usage conditions in the context of standards development are clarified. After further review of those sequences, including initial results of AVC anchor encoding, three additional sequences were selected for the class B (1080 HD) category. It was further decided to remove the previous sequences (same as class A UHD but downsampled, which did not seem to give relevant additional information) from the test set, and retain the corresponding sequences only fo the class A case. Further, for classes A and C, the sequences “Crowd Run” and “Mobisode1” were removed, because they seemed to be redundant in terms of noise and scene characteristics with “Park Joy” and “Mobisode2”, respectively. It is out of question, however, though the current set of sequences may be appropriate for the ongoing Call for Evidence for an initial guess on existence of improved compression technology, it is certainly not good enough for a CfP and following standards development. In particular, material from state-of-the-art 720/60p and 1080/60p cameras, as well as more diversity in the category A (UHD) would be needed.
After the testing cases are settled, the Call for Evidence was issued with responses expected for July. This includes description of the expert viewing methodology which will be applied in London.
6 input documents were registered that present or discuss technical approaches of improved-compression video coding (m16372, m16438, m16448, m16451, m16473, m16479). Comments are included in the list below.
m16372
|
High Definition Test Sequences for High-Performance Video Coding (HVC)
Documentation how the test sequences were produced (including cropping and downsampling)
|
Tomonobu Yoshino
Sei Naito
Shigeyuki Sakazawa
|
m16375
|
AVC Anchor Streams for Evaluation of High-Performance Video Coding (HVC)
Documentation how the AVC anchors were encoded. Some initial findings: QP22 may be too close to transparent for UHD cases; People on Street 5 sec may be too short; motion range may be too low for Park Joy; for WQVGA, difficult to see coding artifacts in case of Keiba, coding artifacts difficult to see in some cases due to high speed of global motion. The two Mobisode sequences may be redundant in terms of their properties.
|
Keiichi Chono
Hirofumi Aoki
Junji Tajime
Yuzo Senda
|
m16438
|
Requests on Coding conditions in Call for Evidence on High-Performance Video Coding (HVC)
Significant reduction of rate if motion vector range is increased for Keiba. Same may be true for Park Joy.
Further discussion on settings after visual review of the results.
|
Yuriy A. Reznik
Ravi K. Chivukula
|
m16448
|
Information on new test material for HVC study
COSME 1920x1080/24p 4:4:4 sequence, provided under clear copyright conditions for development of standards; to be reviewed during the week.
|
Tomoyuki Yamamoto
Tomohiro Ikai
|
m16462
|
Performance Evaluation of Spatially Adaptive Macroblock Size Selection Scheme for HVC Test Sequences
Refers to previous contribution M16082 where fixed (larger) MB size was used. Now, local update of MB size within frames. “Control unit” 64x64, which can be subdivided into 32x32 or 16x16 MBs. Tested on UHD (2560x1600) test sequences from CfE. Most effect in Park Joy (approx. 30% average bitrate reduction). Other sequences small or no improvement. Only IPPP investigated. Results were made with different search range for each of the different MB sizes.
Note: Current document is without PSNR figures. Will be updated.
|
Steffen Kamp
Mathias Wien
|
m16463
|
On design of transforms for high-resolution / high-performance video coding
To be presented later during the week.
|
Steffen Kamp
Mathias Wien
|
m16472
|
Analysis on partition and transform selection in the context of extended block sizes
MB sizes up to 64x64; transform size up to 16x16 (also 16x8 and 8x16; in some cases 16x1 is used). Larger partitions are typically more often used when QP becomes higher. Based on statistics, certain transform sizes are disabled in combination with certain MB sizes. Only IPPP coding. 0.3% average bitrate reduction, but is is claimed that complexity is reduced.
|
Yoshihisa Yamada
Yoshiaki Kato
Kohtaro Asai
Tokumichi Murakami
|
m16473
|
Additional coding performance evaluation of extended MB size
Extended MB size to maximum 32x32, motion partitioning 32x16 and 16x32 additionally. P picture only. For UHD sequences, relevant gain is observed only for traffic (13%), other sequences low gain, on average around 3-4%. Claimed that 2 out of 4 sequences are film-scan sources that may prevent good inter coding. Gain for other 1080p and 720p sequences higher (10-15% average).
|
Kazuo Sugimoto
Yoshihisa Yamada
Yoshiaki Kato
Kohtaro Asai
Tokumichi Murakami
|
m16479
|
Additional Experimental Result of MVOP with HD Sequences
Motion vector coding with optimum predictor: New results for 720p and 1080p sequences. Decoder decides on optimum motion vector to be used for prediction, but flag is used to signal that the median should be used instead. 5 candidates (3 spatial neighbors, co-located MB in previous frame, median). Template matching is used to determine the optimum. Experiments with only CAVLC, IPPP. Average for 720p 8.7%, for 1080p 7.8% (much less for Crowd Run and Park Joy). Usually, the performance goes up with increased spatial resolution.
|
Jungyoup Yang
Kwanghyun Won
Byeungwoo Jeon
Su Nyeon Kim
|
Output documents:
Dostları ilə paylaş: |
