1. Overview of the Scalable Video Coding Extension of the H
Overview of the Scalable Video Coding Extension of the H.264/AVC Standard
Talk about SVC
Heiko Schwarz, Detlev Marpe, Member, IEEE, and Thomas Wiegand, Member, IEEE
presentation by: Fred Scott
adapted from: Kianoosh Mokhtarian

2. Motivation High heterogeneity among receivers Simulcasting Transcoding
Connection quality
Display resolution
Processing power
Simulcasting
Transcoding
Scalability
Simulcasting - any number of varying quality streams can be sent but at the cost of a higher bit rate
Transcoding - bad magnification artifacts
Besides, spatial and quality scalability come at cost of significant loss in coding efficiency and large increase in decoder complexity.

3. Overview Background Temporal scalability Spatial scalability
Quality scalability
Conclusion
Background: types of scalability, applications, and requirements

4. Background Scalability Applications Temporal Spatial
Quality (fidelity or SNR)
Object-based and region-of-interest
Hybrid
Applications
Encode once, decode with differing quality
Unequal importance + unequal error protection
Player sensitive
Definition of scalability: roughly: parts of the video bitstream can be removed in a way that the resulting …
Object based and ROI scalability: Substreams represent spatially contiguous regions.
Unequal error protection: esp. for the case of unpredictable throughput variations and/or high loss rate - error protection is greatly improved
Player sens - ipod vs sdtv vs hdtv, low delay stream, later dl for full quality on slower connections.

5. Background Requirements for a scalable video coding technique
Similar coding efficiency to single-layer coding
Little increase in decoding complexity
Support of temporal, spatial, quality scalability
Backward compatibility of the base layer
Support of simple bitstream adaptations after encoding
Similar coding efficiency: similarity means: for each substream: 10% to up to 50%, depending on the specific needs of an application.
Non-VCL NAL units: Information that change infrequently.
Video sequence: independently decodable part of a NAL unit bitstream.
IDR is an access unit.

6. Overview Background Temporal scalability Spatial scalability
Quality scalability
Conclusion

7. Temporal Scalability
Enabled by restricting motion-compensated prediction
Already provided by H.264/AVC
Hierarchical prediction structure
Pictures of temporal enhancement layers: typically B- pictures
Group of Pictures (GoP)
modifications to the standards of temporal encoding can lead to scalability
rather than minimizing bitstream, the GOP’s can be arranged around

8. Temporal Scalability: Hierarchical Pred’ Struct’
Dyadic temporal enhancement layers
dyadic - pairs of frames

9. Temporal Scalability: Hierarchical Pred’ Struct’
Non-dyadic case
it was just a special case:
Not only the dyadic case: the example shows the case: susbtreams are of 1/3 and 1/9 of the full frame rate.

10. Temporal Scalability: Hierarchical Pred’ Struct’
Other flexibilities
Multiple reference picture concept of H.264/AVC
Reference picture can be in the same layer as the target frame
Hierarchical prediction structure can be modified over time
Also, prediction structure of the base layer can be arbitrarily modified, e.g., such as for increasing coding efficiency.

11. Temporal Scalability: Hierarchical Pred’ Struct’
Adjusting the structural delay
any bidirectionality can add delays
This is an example of adjusting the structure such that the delay becomes zero, i.e., receiving and decoding order become the same.
Every picture can use only reference pictures that precede it.
It typically decreases coding efficiency.

12. Temporal Scalability: Coding Efficiency
Highly dependent on quantization parameters
Intuitively, higher fidelity for the temporal base layer pictures
How to choose QPs
Expensive rate-distortion analysis
QPT = QP T
High PSNR fluctuations inside a GoP
Subjectively shown to be temporally smooth
Dependence on quantization parameter: exists for both scalable and non- scalable cases.
QP = Quantization Parameters
QP_0: that of the base layer, QP_T: of the temporal layer T.
This equation is tested for a wide range of sequences.
Temporally smooth: without the annoying temporal artifacts.

13. Temporal Scalability: Coding Efficiency
Dyadic hierarchical B-pictures, no delay constraint
Foreman, CIF, 30 Hz.
referring GOP arrangement
Coding efficiency is continuously improved by increasing GoP size.
In comparison to the widely used IBBP coding structure, PSNR gain of more than 1 dB can be obtained for medium bitrates.

14. Temporal Scalability: Coding Efficiency
High-delay test set, CIF 30Hz, 34dB, compared to IPPP

15. Temporal Scalability: Coding Efficiency
Low-delay test set, 365x288, 25-30Hz, 38dB, delay is constrained to be zero compared to IPPP
Low delay test set: video conferencing sequences.
Coding efficiencies are significantly smaller than the previous slide.

16. Temporal Scalability: Conclusion
Typically no negative impact on coding efficiency
But also significant improvement, especially when higher delays are tolerable
Minor losses in coding efficiency are possible when low delay is required
**Temporal scalability does not have any significant impact on coding efficiency, except in a few cases requiring low delays

17. Overview Background Temporal scalability Spatial scalability
Quality scalability
Conclusion

18. Spatial Scalability
Motion-compensated prediction and intra-prediction in each spatial layer, as for single-layer coding
Inter-layer prediction
Same coding order for all layers
multilayered - each one for a specific spatial resolution
interlayer for coding efficiency
Same coding order: to restrict memory requirements and decoder complexity.

19. Spatial Scalability
Motion-compensated prediction and intra-prediction in each spatial layer, as for single-layer coding
Inter-layer prediction
Same coding order for all layers
Access units
Access units: lower layer pictures do not need to be present in all access units.
Combination of spatial and temporal scalabilities.

20. Spatial Scalability: Inter-Layer Prediction
Previous standards
Inter-layer prediction by upsampling the reconstructed samples of the lower layer signal
Prediction signal formed by:
Upsampled lower layer signal
Temporal prediction inside the enhancement layer
Averaging both
Lower layer samples not necessarily the most suitable data for inter-layer prediction
Prediction of macroblock modes and associated motion parameters
Prediction of the residual signal
Two additional inter-layer prediction concepts …
the temporal prediction signal mostly represents a better approximation of the original signal than the upsampled lower layer reconstruction.

21. Spatial Scalability: Inter-Layer Prediction
A new macroblock type signalled by base mode flag
Only a residual signal is transmitted
No intra-prediction mode or motion parameter
If the corresponding block in the reference layer is:
Intra-coded  inter-layer intra prediction
The reconstructed intra-signal of the reference layer is upsampled as a predictor
Inter-coded  inter-layer motion prediction
Partitioning data are upsampled, reference indexes are copied, and motion vectors are scaled up
Derivation is done based on the corresponding data of the co-located 8x8 block in the reference layer:
Done by upsampling/scaling up: partitioning, not indexes, vectors.

22. Spatial Scalability: Inter-Layer Prediction
Inter-layer motion prediction (for a 16x16, 16x8, 8x16, or 8x8 macroblock partition)
Reference indexes are copied
Scaled motion vectors are used as motion vector predictors
Inter-layer residual prediction
Can be used for any inter-coded macroblock, regardless of its base mode flag or inter-layer motion prediction
The residual signal of the reference layer is upsampled as a predictor

23. Spatial Scalability: Inter-Layer Prediction
For a 16x16 macroblock in an enhancement layer:
Inter-layer intra prediction (samples values are predicted) 1
Inter-layer residual prediction
Inter-layer motion prediction (partitioning data, ref. indexes, and motion vectors are derived)
base
mode
flag
A summary
No inter-layer residual prediction
Inter-layer motion prediction (ref. indexes are derived, motion vectors are predicted)
No inter-layer motion prediction

24. Spatial Scalability: Generalizing
Not only dyadic
Enhancement layer may represent only a selected rectangular area of its reference layer picture
Enhancement layer may contain additional parts beyond the borders of its reference layer picture
Tools for spatial scalable coding of interlaced sources
Like previous standards, it is not limited only to the dyadic case.

25. Spatial Scalability: Complexity Constraints
Inter-layer intra-prediction is restricted
Although coding efficiency is improved by generally allowing this prediction mode
Each layer can be decoded by a single motion compensation loop, unlike previous coding standards
Restricted only to macroblocks whose co-located blocks are intra-coded.

26. Spatial Scalability: Coding Efficiency
Comparison to single-layer coding and simulcast
Base/enhancement layer at 352x288 / 704x576
Only the first
frame is
intra-coded
Inter-layer
prediction (ILP):
Intra (I),
motion (M),
residual (R)
“City” was the worst performing case in the test set.

27. Spatial Scalability: Coding Efficiency
Comparison to single-layer coding and simulcast
Base/enhancement layer at 352x288 / 704x576
Only the first
frame is
intra-coded
Inter-layer
prediction (ILP):
Intra (I),
motion (M),
residual (R)
As can be seen, effectiveness of a prediction tool or combination of tools strongly depends on the sequence characteristics.

28. Spatial Scalability: Coding Efficiency
Comparison to single-layer coding and simulcast
Base/enhancement layer at 352x288 / 704x576
Only the first
frame is
intra-coded
Inter-layer
prediction (ILP):
Intra (I),
motion (M),
residual (R)
Also, effectiveness of a prediction tool or combination of tools strongly depends on the prediction structure.
Rate-distortion performance of SVC compared to single-layer coding reduces when moving from a GoP size of 16 to IPPP coding.

29. Spatial Scalability: Coding Efficiency
Comparison of fully featured
SVC “single-loop ILP (I, M, R)”
to scalable profiles of previous
standards “multi-loop ILP (I)”
The gain of multi-loop decoding is often minor, and brings significant decoding complexity.

30. Spatial Scalability: Encoder Control
JSVM software encoder control
Base layer coding parameters are optimized for that layer only
 performance equal to single-layer H.264/AVC
Spatial Scalability: Encoder Control
Joint Scalable Video Model

31. Spatial Scalability: Encoder Control
JSVM software encoder control
Base layer coding parameters are optimized for that layer only
 performance equal to single-layer H.264/AVC
Not necessarily suitable for an efficient enhancement layer coding
Improved multi-layer encoder control
Optimized for both layers

32. Spatial Scalability: Encoder Control
QPenhancement layer = QPbase layer + 4
Hierarchical B-pictures, GoP size = 16
Bit-rate increase relative to single-layer for the same quality is always less than or equal to 10% for both layers

33. Overview Background Temporal scalability Spatial scalability
Quality scalability
Conclusion

34. Quality Scalability
Special case of spatial scalability with identical picture sizes
No upsampling for inter-layer predictions
Inter-layer intra- and residual-prediction are directly performed in transform domain
Different qualities achieved by decreasing quantization step along the layers
Coarse-Grained Scalability (CGS)
A few selected bitrates are supported in the scalable bitstream
Quality scalability becomes less efficient when bitrate difference between CGS layers gets smaller

35. Quality Scalability: MGS
Medium-Grained Scalability (MGS) improves:
Flexibility of the stream
Packet-level quality scalability
Error robustness
Controlling drift propagation
Coding efficiency
Use of more information for temporal prediction

36. Quality Scalability: MGS
MGS: error robustness vs. coding efficiency A B
Various concepts for trading off enhancement layer coding efficiency and drift for packet-based quality scalable coding.
(a) Base layer only control.
(b) Enhancement layer only control.
(c) Two-loop control. (d) Key picture concept of SVC for hierarchical prediction structures, where key pictures are marked by the hatched boxes. C D

37. Quality Scalability: MGS
MGS: error robustness vs. coding efficiency
Pictures of the coarsest temporal layer are transmitted as key pictures
Only for them the base layer picture needs to be present in decoding buffer
Re-synchronization points for controlling drift propagation
All other pictures use the highest available quality picture of the reference frames for motion compensation
High coding efficiency

38. Quality Scalability: Encoding, Extracting
Encoder does not known what quality will be available in the decoder
Better to use highest quality references
Should not be mistaken with open-loop coding
Bitstream extraction
based on priority identifier of NAL units assigned by encoder

39. Quality Scalability: Coding Efficiency
BL-/EL-only control: motion compensation loop is closed at the base/enhancement layer
2-loop control: one motion compensation loop in each layer
adapt. BL/EL control: use of key pictures

40. Quality Scalability: Coding Efficiency
MGS vs. CGS

41. SVC encoder structure example

42. Overview Background Temporal scalability Spatial scalability
Quality scalability
Conclusion

43. Conclusion SVC outperforms previous scalable video coding standards
Hierarchical Structures
Temporal and Spatial
Inter-layer and Intra-layer prediction
Medium Grain Scalability (MGS)

44. References
H. Schwarz, D. Marpe, and T. Wiegand, “Overview of the scalable video coding extension of the H.264/AVC standard,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 17, no. 9, pp. 1103– 1120, September 2007.
T.Wiegand, G. Sullivan, J. Reichel, H. Schwarz, and M.Wien, "Joint Draft ITU-T Rec. H.264 | ISO/IEC / Amd.3 Scalable video coding," Joint Video Team, Doc. JVT-X201, July 2007.
H. Kirchhoffer, H. Schwarz, and T. Wiegand, "CE1: Simplified FGS," Joint Video Team, Doc. JVT-W090, April 2007.