Download as pdf or txt
Download as pdf or txt
You are on page 1of 5

FAST PARTITIONING DECISION STRATEGIES

FOR THE UPCOMING VERSATILE VIDEO CODING (VVC) STANDARD

Adam Wieckowski*, Jackie Ma*, Heiko Schwarz*†, Detlev Marpe*, Thomas Wiegand*‡
* † ‡
Fraunhofer HHI Freie Universität Berlin Technische Universität Berlin
Video Coding & Analytics Inst. of Computer Science Media Technology
Berlin, Germany Berlin, Germany Berlin, Germany
ABSTRACT illustrates the different split modes that are supported in the
current Test Model VTM-3.0 of Versatile Video Coding.
One of the key technologies in the upcoming Versatile Video The combination of QTBT and the ternary split is called
Coding (VVC) standard is the flexible partitioning. QTBTT and replaces the quad-tree structure used in HEVC.
Compared to HEVC, partitioning alone accounts for about In general, the support of additional partitioning modes
8.5% bitrate savings. The added flexibility comes at a cost of typically increases the achievable coding efficiency for a
large search space expansion. video coding design. However, it also increases the encoder
This paper gives an overview of fast encoder strategies complexity, since an encoder has to test more splitting
for selecting the block partitioning in the Test Model VTM- options to select the appropriate partitioning for a picture
3.0 for VVC. In total, 13 approaches for speeding up the region. The first test model of VVC, VTM-1.0 [5], employs
block partitioning selection in an encoder are described; 8 of the described QTBTT partitioning, but despite of that it
these techniques were developed by the authors. 10 of the 13 supported a very similar set of coding tools as HEVC. The
methods are used in the common test conditions of the Joint more flexible partitioning yields average bit rate savings of
Video Experts Team (JVET); these 10 methods achieve an about 8.5% relative to HEVC [6].
overall speed-up factor of 7 at a bit rate increase of only 1.1%. In modern video encoders, as the VTM-3.0 reference
By enabling 12 approaches, a speed up factor of over 8.5 is encoder, the coding modes for a block of samples is selected
obtained at a bit rate increase of about 1.5%. by minimizing a Lagrangian rate-distortion (RD) cost [7],
J = D + λ ⋅ R, (1)
Index Terms— VVC, QTBT, speed-up, partitioning.
where 𝐷 represents the resulting SSD distortion between
original and reconstructed samples, 𝑅 specifies the number of
1. INTRODUCTION
bits required for coding the block in a certain coding mode,
Similar as its predecessor HEVC, the upcoming video coding and 𝜆 is a constant Lagrange multiplier. When deciding
standard Versatile Video Coding (VVC) follows the concept between two modes, their RD costs J are compared and the
of block-based hybrid video coding. While HEVC [1] only mode with lower cost is chosen. The partitioning search is
supports a block partitioning using quadtrees [2], the current based on the G-BFOS algorithm [2], in which a cost Js of a
Test Model (VTM-3.0) for VVC uses more sophisticated and split mode with N sub-blocks is defined as
diverse split modes. In fact, the quad tree (QT) was first Js = λ ⋅ Rs + ∑k Jk. (2)
replaced by the quad-tree plus binary tree (QTBT) structure The split cost Js is calculated from the number of bits required
[3]. QTBT extends the quad-split by an additional binary-tree for signaling the split (Rs) and the sum of the RD costs for the
(BT) split in vertical or horizontal direction to allow a more resulting sub-blocks (Jk), which are given by (1) or (2). The
flexible partitioning of the video pictures. This QTBT estimation of the optimal sub-block RD costs Jk is often a
structure was then further extended by an additional ternary very time consuming process. In a full encoder search, all
split [4], which splits a block into three aligned blocks with possible modes and their coding costs J are calculated.
the center block being twice as large as the outside blocks. A fast encoder search strategy could predict that a cost
Similarly to the binary split, the ternary-tree (TT) split is for a certain mode would not reduce the current minimum
supported in both horizontal and vertical direction. Figure 1 cost J and, then, the full evaluation process of the specific
mode could be skipped. This is particularly important for split
modes, as the estimation of Jk requires the evaluation of all
possible sub-block partitionings.
In the VTM-3.0 reference software, the coding modes
Figure 1: Partitioning modes in VTM-3.0. Quad split, are tested sequentially. For any given block, the split modes
binary splits, and ternary splits. are tested after all non-split modes. Among the split modes,

978-1-5386-6249-6/19/$31.00 ©2019 IEEE 4130 ICIP 2019


the binary splits are tested first, and the quad split is tested as 3.2. BT based QT split prohibition
the last mode. In this paper, we describe several methods by
which unlikely modes are excluded from the mode decision This rule can skip the evaluation of the quad-split, if both
process in order to decrease the encoder run time. binary splits were evaluated and did not reduce the coding
cost. The rule also applies if a binary split was chosen for the
2. RELATED WORK current-block, but none of the sub-blocks is further split.

For the HEVC quad-tree partitioning, numerous fast 3.3. Adaptive QT search restriction
encoding strategies have been proposed in the literature [2,8-
16]. A minimal and maximal number of consecutive QT-splits is
Some of these methods are directly related to (1) and (2). adaptively calculated from already estimated neighboring
For example methods presented in [2] and [8] discard any blocks. A QT split is enforced if the current depth is below
further splitting after a specific cost threshold is reached for the range and a QT split is prohibited if the current depth
the coding of a current block. More precisely, the threshold exceeds the range, respectively.
used in [2] is based on the sum of absolute transform
coefficients, whereas the threshold proposed in [8] is 3.4. Depth-restricted Skip mode condition
obtained by weighting the split signaling rate Rs and the
current non-split cost with a context dependent parameter.
In [18], a probabilistic model for predicting the cost of This speed-up is similar to the method described in 3.1, but
splitting is developed for QTBT; this approach is similar to its application is restricted to nodes at predefined depths of
the simple predictor suggested in [8]. This strategy allows to the BTT (i.e. QTBTT without QT), thus only skipping the
skip the evaluation of about 50% of the supported partitioning search of relatively small portions of the partitioning space.
modes, but results in a 1% bit rate increase (for the JEM
software) [19]. A detailed review of further methods is 3.5. Block search history heuristics
beyond the scope of this paper; the interested reader is
referred to overview in [8]. This method does not control the decision to perform an RD
With QTBTT for VVC [17] being a very recent search for a specific split, but rather controls the RD search
partitioning approach still under development, at the time of for non-split modes exploiting partitioning redundancies.
writing, the authors are not aware of any fast encoder strategy With QTBTT, a specific coding block might results from
specifically designed for the QTBTT partitioning. different partitioning configurations – e.g. 64×64 block might
results from a quad split at CTU (128×128) level, or a binary
3. FAST ENCODER STRATEGIES vertical split followed by a binary horizontal split. In this
method, when a specific coding block is first evaluated, the
In the following, we briefly describe different methods for RD search results are stored. Every time this block is
accelerating the partitioning search in VVC. Table 1 contains evaluated again in different partitioning modes, the stored
a summary of all discussed approaches. The approaches information is used to form a set of heuristic rules governing
discussed in sections 3.1.–3.5. were initially developed for the selection of non-split modes used in the RD search.
the JEM [19] software, and ported into the initial version
VVC VTM-1.0. The strategies described in sections 3.6.– 3.6. Direct block re-usage
3.13. were introduced by the authors in various standard
contributions. This method is similar to the approach described in 3.5. In
While additional speed-ups were introduced with the contrast to 3.5., the best result from the first evaluation is
initial adoption of QTBT [20,21], some of them are not directly reused instead of forming a set of heuristic rules. This
related to selecting the block partitioning and are thus not method can be applied when the causal neighborhood of a
discussed in this paper. Such speed-ups include an improved block is in the exact same state as during the first evaluation
selection of merge modes and a fast intra-in-inter decision. – i.e. they have the same CABAC engine state (except for the
partitioning-related contexts) as well as all neighboring
3.1. Early Skip-based termination condition reconstructed samples used for prediction are the same.

This rule was introduced in the HEVC reference software 3.7. QT first RD search
[22]. If the Skip mode is the best non-split mode for a current
block (in terms of the RD cost), then no split mode is tested. This method changes the order of split modes. If activated,
The rule works well for HEVC, but yields a large loss in the quad split will be tested before the binary and ternary
coding efficiency for VVC (as shown in Table 1). splits. The binary and ternary splits are then only tested if no
further quad split is selected by the RD search in any of the

4131
quadrants. To not interfere with the speed-up described in 3.10. Skip history based speed-up
3.2., the change in search order and the application of this
heuristic is conditioned on the quad-tree depth of the This rule is based on the skip mode. However, in contrast to
neighboring blocks, which indicates the probability of a 3.1. and 3.4., it also considers the history of past best coding
quad-tree split being applied to the current block. modes. For non-intra slices no split modes are being tested, if
for the past two consecutive coding levels (i.e. the current
3.8. Topology-based TT split prohibition block and the block for which the current block is a sub-
block) the skip mode was the best coding option.
In general, the ternary split isolates the middle of a block in a 3.11. Residual-based TT split prohibition
larger block and the edges in smaller blocks. Thereby it
produces two elongated rectangular blocks at the boundary. A ternary split is not tested for a block at the current level if
This method evaluates the topology resulting from evaluating the best non-split coding mode has no residual and the
binary splits to decide the viability of testing parallel ternary parallel binary split (tested before the ternary split) did not
splits. If the results of binary splits do not reflect this reduce the RD cost.
property, i.e., neither of the binary split sub-blocks is further
split in the same direction, the RD search for the ternary split 3.12. Content-based gradient speed-up
is skipped.
This rule is based on the original signal. Splits in horizontal
direction are not tested if the gradient of the original block in
3.9. Split cost prediction vertical direction is greater than the gradient in horizontal
direction and the gradient along the diagonal. Likewise, splits
The idea behind this method [8] is to predict the resulting RD in vertical direction are not tested if the gradient in horizontal
cost J of a split without an extensive RD search, using a direction is greater than the gradient in vertical direction and
simple expression Jp. The predictor uses the known best non- the gradient along the diagonal. Further, a threshold controls
split cost J0 and weights it with a configurable factor f. With the strictness of this rule.
Rs being the rate required for signaling the split mode, the
predicted cost is derived by (see [8]) 3.13. Heuristic intra leaf cut
Jp = λ ⋅ Rs + J0 / f. (3)
No further split modes are tested for blocks with size smaller
If Jp > J0, then the RD search for the specific split is skipped. than 32×32 in all-intra configuration if the current best coding
The factor f is derived from the quantization parameter QP; it unit has no residual and the current binary tree depth is strictly
is equal to 1.1 for QPs larger than 30, and 1.075 otherwise. greater than one. This rule is only applied to luma blocks.

Table 1: List of reviewed rules and their performance impact for random-access (RA) and all-intra (AI) configurations.
The speed-up impact is tested for a configuration without any partitioning speed-ups (“on”) and a
configuration with all other speed-ups included in CTC enabled (“off”). Last two rows show the combined
performance of all methods enabled in CTC, and all described methods. Inactive rules are marked cursive.
RA on RA off AI on AI off
Name Description Adoption BD-Y EncT BD-Y EncT BD-Y EncT BD-Y EncT
ECU 3.1. [22] 3.35% 37% 3.28% 80% 0.00% 100% 0.00% 100%
QT_FROM_BT 3.2. [21] -0.01% 67% 0.01% 94% 0.01% 82% 0.00% 97%
ADAPT_QTD 3.3. [26] 0.00% 94% 0.04% 100% 0.01% 97% 0.00% 100%
SKIPD 3.4. [25] 0.11% 77% 0.11% 95% 0.00% 100% 0.00% 100%
REUSE_HEU 3.5. [27] 0.35% 68% 0.22% 70% 0.00% 100% 0.00% 100%
JEM 3.2. – 3.5. 0.47% 34% 0.38% 63% 0.02% 81% -0.02% 98%
REUSE_DIRECT 3.6. [28] -0.01% 90% 0.02% 95% 0.00% 100% 0.00% 100%
BT_FROM_QT 3.7. [29] 0.09% 96% 0.04% 88% 0.06% 92% 0.05% 90%
TT_FROM_BT 3.8. [30] 0.39% 89% 0.37% 81% 0.24% 85% 0.24% 78%
SPLIT_COST 3.9. [28] 0.02% 49% 0.06% 84% 0.02% 62% 0.05% 75%
SKIP_HIST 3.10. [30] 0.07% 58% 0.07% 97% 0.00% 100% 0.00% 100%
BTT_CBF 3.11. [30] 0.08% 59% 0.11% 85% 0.00% 84% -0.01% 99%
GRAD 3.12. [31] 0.45% 85% 0.44% 81% 0.23% 86% 0.26% 87%
IS_PLANAR 3.13. [31] 0.00% 100% 0.00% 100% 0.09% 90% 0.09% 87%
NONJEM-CTC 3.6. – 3.11. 0.71% 23% 0.62% 39% 0.35% 41% 0.31% 48%
NONJEM-ALL 3.6. – 3.13. 1.11% 19% 1.06% 32% 0.67% 35% 0.66% 41%
CTC 3.2. – 3.11. 1.09% 14% 0.33% 41%
ALL 3.2. – 3.13. 1.54% 12% 0.68% 34%

4132
4. RESULTS rules are based on the same assumption. During the encoder
search, at some point the encoder will decide for a block that
The individual and combined results for all described the optimal encoding is not to code any residual and possibly
methods are presented in Table 1. The results are based on use the simplest prediction – which is the Skip mode. This
the test model VTM-3.0 in two configurations (for the “on” might be attributed to the fact that this simple prediction is
and “off” tests). The table shows the reported luma BD rate almost perfect and no further refinement is necessary. It
[6] change and the difference of run times (using geometric might be also caused by the residual containing essential high
means as specified for reporting in common test conditions frequencies which are very costly to encode. In the latter case,
CTC [23]) over all CTC sequences for each described speed- terminating the encoder search would degrade the coding
up. The table is divided into three parts. The first parts shows performance significantly. This is the reason for the poor
the results of the speed-up rules 3.2. – 3.5., which were results obtained with the ECU method.
inherited from [19]. A combined result for those methods is All of the other aforementioned methods use some form
denoted by JEM. of side-condition to confirm early termination. SKIPD will
The second part of the table shows the speed-ups only terminate if the remaining search space to be searched
developed by the authors 3.6. – 3.13. and incorporated into by the G-BDOF algorithm [2] is relatively small. SKIP_HIST
the VVC test model [24] in various contributions. This part is checks if the decision is repeated at two consecutive levels of
summarized showing the combined performance of the the partitioning hierarchy, and BTT_CBF only skips the RD
speed-ups for all of the proposed speed-ups (NONJEM search of a TT split if the parallel BT split did not provide a
ALL) and only the speed-ups 3.6. – 3.11., which are part of coding improvement. The method SPLIT_COST compares
the current common test conditions [23] (NONJEM-CTC). the current best non-split coding cost against the known rate
In the last part of the table, the results for all combined for split signaling. In this way, the condition is based on the
speed-ups are denoted as ALL (excluding ECU due to poor actual coding cost of the best candidate and as such is more
performance). Speed-ups currently active in the common test flexible and adaptive to the true coding cost.
conditions [23] are presented as CTC. The rules SPLIT_COST and BTT_CBF keep a
Each speed-up is tested in two configurations. First in an considerable part of their performance in the “off” test,
“on” configuration and second in an “off” configuration. In meaning that at least partially they introduce unique
the “on” configuration, all discussed methods but the tested termination strategies that are not encompassed by any other
speed-up are disabled. In the “off” configuration, the speed- speed-ups. In contrast, SKIPD and SKIP_HIST lose most of
up is tested against the CTC configuration, i.e. all speed-ups their performance in the “off” test indicating that most of the
3.2.-3.11. are on except for the speed-up that is to be tested. early terminations they introduce can be expressed using one
Thus, the “off” tests show how the specified method performs of the remaining rules.
in the context of already heavily optimized partitioning The speed-ups REUSE_HEU and GRAD perform very
search. The differences in both measurements are discussed similar in both “on” and “off” test, implicating that they are
in the next section. Due to high computational complexity for more or less independent from other methods.
the “on” tests, the results are shown for the first 49-frames. Notably, the rules BT_FROM_QT and TT_FROM_BT,
The overall performance when speed-ups are activated in show better performance in the “off” tests than in the “on”
the common test conditions for the random-access tests. This indicates these methods have less interaction with
configuration shows a 7× speed-up with a luma BD-rate loss other speed-ups.
of 1.09%. When all speed-ups are activated the speed-up can
be increased to 8.5× with a luma BD-rate loss of about 1.5%. 6. CONCLUSION
The contributions by the authors to the partitioning
encoder decisions on top of the pre-existing JEM rules The upcoming video coding standard VVC utilizes a very
accounts for a 2.5× speed-up at a 0.62% luma BD-rate loss. powerful partitioning scheme. It allows to isolate complex
When considered independently (“on” configuration), these patterns into simple rectangular shapes. This flexibility
rules can speed-up the partitioning by a factor of over 4 for a creates a very large search space, which needs to be processed
BD-rate loss of 0.71%. efficiently. We reviewed the search rules used in the current
reference software encoder [24] with an emphasis on the
5. DISCUSSION speed-ups contributed by the authors.
The overall bit rate savings introduced by the new
In the results presented in Table 1 it is apparent that most of partitioning structure is about 8.5% [5] (with some of the
the presented rules show better results in the “on” case than described speed-ups already included). Using the current
in the “off” test. This can be mostly attributed to fact, that in algorithm in the random-access configuration, the encoder
the “on” test, the tested search space is much larger, thus can skip searching through 86% of the partitioning search
containing more unnecessary tests. space with a bit rate increase of only 1.09% (about 13% of
This phenomenon is the strongest for the rules ECU, the overall partitioning gain), thus providing a 7× encoder
SKIPD, SPLIT_COST, SKIP_HIST and BTT_CBF. These speed-up.

4133
7. REFERENCES (ICME): Seattle, USA, July 11-15, 2016, Seattle, WA, USA,
2016.
[1] G. J. Sullivan, J.-R. Ohm, W.-J. Han, and T. Wiegand, [17] B. Bross, J. Chen, and S. Liu, Versatile Video Coding (Draft
“Overview of the High Efficiency Video Coding (HEVC) 3), document JVET-L1001, Joint Video Experts Team
Standard,” IEEE Transactions on Circuits and Systems for (JVET), Oct. 2018.
Video Technology, vol. 22, no. 12, pp. 1649–1668, 2012. [18] Z. Wang, S. Wang, J. Zhang, S. Wang, and S. Ma,
[2] D. Marpe et al., “Video Compression Using Nested Quadtree “Probabilistic Decision Based Block Partitioning for Future
Structures, Leaf Merging, and Improved Techniques for Video Coding,” IEEE transactions on image processing: a
Motion Representation and Entropy Coding,” IEEE publication of the IEEE Signal Processing Society, 2017.
Transactions on Circuits and Systems for Video Technology, [19] X. Li, and K. Suehring, JVET AHG report: JEM software
vol. 20, no. 12, pp. 1676–1687, 2010. development (AHG3), document JVET-J0003, Joint Video
[3] J. An, Y.-W. Chen, K. Zhang, H. Huang, Y.-W. Huang, and S. Experts Team (JVET), Apr. 2018.
Lei, Block partitioning structure for next generation video [20] J. An, H. Huang, K. Zhang, Y.-W. Huang, and S. Lei,
coding, ITU-T SG16 Doc. COM16–C966, Sep. 2015. Quadtree plus binary tree structure integration with JEM
[4] X. Li, H.-C. Chuang, J. Chen, M. Karczewicz, L. Zhang, X. tools, document JVET-B0023, Joint Video Exploration Team
Zhao, and A. Said, Mult-Type-Tree, document D0117, Joint (JVET), Feb. 2016.
Video Exploration Team (JVET), Oct. 2016. [21] H. Huang, K. Zhang, Y.-W. Huang, and S. Lei, EE2.1:
[5] F. Bossen, X. Li, and K. Suehring, JVET AHG report: Test Quadtree plus binary tree structure integration with JEM
model software development (AHG3), document JVET- tools, document JVET-C0024, Joint Video Exploration Team
K0003, Joint Video Experts Team (JVET), Jul. 2018. (JVET), May 2016.
[6] G. Bjøntegaard, Calculation of average PSNR differences [22] R. H. Gweon, Y.-L. Lee, and J. Lim, Early Termination of CU
between RD-curves, Technical Report VCEG-M33, ITU-T Encoding to Reduce HEVC Complexity, document JCTVC-
SG16/Q6, Austin, Texas, USA, 2001. F045, Joint Collaborative Team on Video Coding (JCT-VC),
[7] G. J. Sullivan and T. Wiegand, “Rate-Distortion Optimization Jul. 2011.
for Video Compression,” IEEE Signal Processing Magazine, [23] F. Bossen, J. Boyce, X. Li, V. Seregin, and K. Suehring, JVET
vol. 15, no. 6, pp. 74–90, 1998. common test conditions and software reference configurations
[8] A. Wieckowski, J. Ma, H. Schwarz, D. Marpe, and T. for SDR video, document JVET-L1010, Joint Video Experts
Wiegand, “Recursive partitioning search space pruning using Team (JVET), Oct. 2018.
split cost prediction,” in 2019 Data Compression Conference [24] F. Bossen, X. Li, and K. Suehring, JVET AHG report: Test
(DCC), Snowbird, UT, 2019. model software development (AHG3), document JVET-
[9] G. Correa, et al., “Fast HEVC Encoding Decisions Using Data M0003, Joint Video Experts Team (JVET). Jan. 2019.
Mining,” Circuits and Systems for Video Technology, IEEE [25] Y. Yamamoto, T. Ikai, and Y. Yasugi, AHG5: Improved fast
Transactions on, vol. 25, no. 4, pp. 660–673, 2015. encoding setting, document JVET-E0023, Joint Video
[10] J. Chen and L. Yu, “Effective HEVC intra coding unit size Exploration Team (JVET), Jan. 2017.
decision based on online progressive Bayesian classification,” [26] J Chen, Y Chen, M Karczewicz, X Li, H Liu, L Zhang, and X
in 2016 IEEE International Conference on Multimedia & Expo Zhao, Coding tools investigation for next generation video
(ICME): Seattle, USA, July 11-15, 2016, Seattle, WA, USA, coding, ITU-T SG16 Doc. COM16-C806-E, Sep. 2015.
2016. [27] H. Huang, S. Liu, Y.-W. Huang, C.-Y. Chen, and S. Lei,
[11] Y. Liu, W. Tian, Y. Li and H. Yang, “Fast CU Termination AHG5: Speed-up for JEM-3.1, document JVET-D0077, Joint
Algorithm with AdaBoost Classifier in HEVC Encoder,” Video Exploration Team (JVET), Oct. 2016.
IEICE Trans. Inf. & Syst., vol. E101.D, no. 9, pp. 2220–2223, [28] J. Ma, A. Wieckowski, H. Schwarz, D. Marpe, T. Wiegand, F.
2018. Le Léannec, and T. Poirer, Non-CE1: Proposal for a
[12] M. Tang et al., “Accelerating HEVC Encoding Using Early- partitioning method by Franhofer HHI and Technicolor,
Split,” Signal Processing Letters, IEEE, vol. 25, no. 2, pp. 209– document JVET-K0220, Joint Video Experts Team (JVET),
213, 2018. Jul. 2018.
[13] Y. Zhang et al., “Machine learning-based coding unit depth [29] A. Wieckowski, T. Hinz, V. George, J. Brandenburg, J. Ma, S.
decisions for flexible complexity allocation in high efficiency De-Luxán-Hernández, B. Bross, H. Schwarz, D. Marpe, and T.
video coding,” (eng), IEEE transactions on image processing: Wiegand, AHG10: Updated NextSoftware as an alternative
a publication of the IEEE Signal Processing Society, vol. 24, implementation of the Joint Exploration Model (JEM),
no. 7, pp. 2225–2238, 2015. document JVET-I0034, Joint Video Experts Team (JVET),
[14] S. Zhou, Z. Ye, and Y. Wang, “Fast HEVC CU/PU mode Jan. 2018.
decision based on ANN and texture analysis,” in 2016 Sixth [30] A. Wieckowski, T. Hinz, B. Bross, T. Nguyen, J. Ma, K.
International Conference on Image Processing Theory, Tools Suehring, H. Schwarz, D. Marpe, and T. Wiegand,
and Applications: IPTA 2016, Oulu, Finland, December 2016, NextSoftware as Test Software, document JVET-J0095, Joint
Oulu, Finland, 2016. Video Experts Team (JVET), Apr. 2018.
[15] H. Amer, A. Rashwan and E. Yang, “Fully Connected [31] A. Wieckowski, T. Hinz, V. George, J. Brandenburg, J. Ma, B.
Network for HEVC CU Split Decision equipped with Bross, H. Schwarz, D. Marpe, and T. Wiegand, NextSoftware:
Laplacian Transparent Composite Model,” in 2018 Picture An alternative implementation of the Joint Exploration Model
Coding Symposium (PCS), San Francisco, CA, 2018. (JEM), document H0084, Joint Video Exploration Team
[16] J. Chen and L. Yu, “Effective HEVC intra coding unit size (JVET), Oct. 2017.
decision based on online progressive Bayesian classification,”
in 2016 IEEE International Conference on Multimedia & Expo

4134

You might also like