|
It is certainly good for MPEG to look into possible future progress in video coding as this is a major core area of MPEG expertise. Two contributions were reviewed in the video group, which show the potential of new applications (new display types with larger ranges of color and brightness dynamics, higher resolution) appearing (m15409), as well as the potential that higher compression could be achieved than currently by using AVC (m15375).
As a first action on this, a brainstorming discussion was organized on Wednesday afternoon. In this context, various points were discussed such as
-
New market segments;
-
Competition of MPEG standards vs. proprietary solutions, which may have the benefit of shorter development times and full "hands on” control of the entire encoder/decoder chain, and in this context a) relationship with RVC, which would enable a more “granular” than “monolithic” standards development and b) potential of defining in a normative way components that were traditionally regarded as non-normative in video, such as encoders and pre/post-processing.
In general, it is clear that higher compression is still of significant interest, considering the fact that the number of pixels to be compressed is continuously growing (HD and beyond, higher dynamic ranges, multiview). Furthermore, existing standards such as AVC were mainly optimized for low resolutions (CIF, QCIF) and low rates. As a result from this meeting, a document entitled “First Thoughts on New Challenges in Video Coding Standardization” (N9785) was issued which summarizes the main discussion points. For Hannover, it is planned to run a follow-up brainstorming workshop (AHG installed to prepare for this), soliciting contributions on
-
Is there need for higher compression efficiency? How mature are technical solutions?
-
Studies about performance of our standards in new application domains (including new display types for high resolution, high dynamic range etc.)
-
Studies about application domains where proprietary solutions might be considered as better (including reasons why this is the case)
-
What are criteria to evaluate performance in dedicated application domains?
-
What could be limits of using RVC for extending or developing standards?
-
Is it possible and useful to define performance indicators for encoders, delivery and decoders, would normative encoders be beneficial for certain application domains?
Documents reviewed
m15409
|
Additional results on magnitude-dependent adaptive motion vector coding
In principle same method as proposed by last meeting (quarter pel only for small motion size, half pel up to medium motion size, then full pel).Test with 4:2:0 (high profile) and 4:4:4 predictive profile. Test sequences Cosme, Plane and Crowd Run.
It is noticed that gain primarily comes at medium data rates.
Seems to have most effect in case of global motion blur
For Plane, gain is seen at low rates / qualities (33-37 dB), but no gain or even loss at high rates / qualities (40 dB). For Crowd Run, no gain or loss is observed.
Not just useful for 4:4:4, but also for 4:2:0. The "switching point" would need to be adapted. For plane and 4:4:4 encoding, BD saving is up to 8%, but for other sequences almost zero. Certainly, average savings not sufficient yet for the desirable gain.
Note: Some of the 4:2:0 Plane BD results may not be valid due to crossover of PSNR curves
Hint: Next-generation video sources may require new methods in video coding (careful study needed) and more rate saving.
|
Shun-ichi Sekiguchi
Shuichi Yamagishi
Kenji Otoi
Yoshihisa Yamada
|
m15375
|
Improving AVC compression performance by template matching with decoder-side motion vector derivation
Motion search at decoder side, small search range, reference is an “L shaped region” around current block. Follow-up of a contribution given in Shenzhen, now using a) multiple reference pictures b) averaging of several prediction templates. Only IPPP structure currently used. Usage of TM is signalled as MB mode, encoding by CABAC. Average bit rate saving is 8.6 % average for whole set of CIF and 720p sequences; for 720p alone 11.8 %, for CIF 4.5 %. Main gain at higher rate. Maximum gain for City, 19.2% on average, still most relevant at highest rate point..
Questions:
- Where do the gains come from? Multihypothesis or saving of motion vectors? In fact, multihypothesis would not make sense, if motion vectors for all of the hypotheses would need to be transmitted.
- What impact on complexity? Due to small search ranges around motion predictor, this appears manageable. Even no full search is necessary.
- Would gain further increase with increased search range? Not significantly.
- Increase of memory bandwidth? Due to small search range, this is not of dramatic relevance.
|
Steffen Kamp
Mathias Wien
|
Output document:
Dostları ilə paylaş: |
