These are the slides of my presentation at SPIE Optics + Photonics 2014 Applications of Digital Image Processing XXXVII. The paper itself can be downloaded from SPIE Digital Library. For people in hurry, a pre-print version is available at: http://infoscience.epfl.ch/record/200925?ln=en
Using Grammatical Signals Suitable to Patterns of Idea Development
Comparison of compression efficiency between HEVC and VP9 based on subjective assessments
1. Applications of Digital Image Processing XXXVII
SPIE Optical Engineering + Applications, San Diego, CA, USA, August 18-21, 2014
Martin Rerabek and Touradj Ebrahimi
martin.rerabek@epfl.ch
touradj.ebrahimi@epfl.ch
MMSPG, EPFL, Switzerland
Comparison of compression efficiency
between HEVC and VP9
based on subjective assessments
2. Outline
Introduction
Codecs settings
Dataset
Test methodology
Test equipment and environment
Results
Conclusion
19/8/2014
Applications of Digital Image Processing XXXVII
SPIE Optical Engineering + Applications, San Diego, CA, USA
2
3. Video compression standards for UHDTV
Motivation
– Increased UHDTV content consumption
• Broadcasting, internet streaming
– Video acquisition and display technologies
evolve faster than network capabilities
– Demand for more efficient compression
H264/MPEG-4 Part 10 AVC
H.265/MPEG-H Part 2 HEVC
– Recently released
WebM data format
– VP9 – recently announced
Goals
– Better coding efficiency at higher resolutions
– Suitability to a wide variety of applications
19/8/2014
Applications of Digital Image Processing XXXVII
SPIE Optical Engineering + Applications, San Diego, CA, USA
3
4. HEVC vs VP9 vs AVC comparisons
Can be reliably compared by means of subjective tests
Next-generation encoders
– better compression efficiency is expected for resolutions beyond HDTV
No official subjective evaluations
of HEVC compared to VP9
This paper provides results
of subjective evaluation
on 4K/QFHD video content
– Professional high-performance
4K/QFHD LCD reference monitor
19/8/2014
Applications of Digital Image Processing XXXVII
SPIE Optical Engineering + Applications, San Diego, CA, USA
4
6. Encoding
Standard codecs
– AVC (JM 18.6)
– HEVC (HM 15.0)
Configuration
– Random Access
– GOP size: 8 pictures
– Intra Period: 1s (broadcast scenario)
Hierarchical B-pictures
QP increase of 1 between
each Temporal Level
– Coding Order: 0 8 4 2 1 3 6 5 7
19/8/2014
Applications of Digital Image Processing XXXVII
SPIE Optical Engineering + Applications, San Diego, CA, USA
6
7. Encoding
VP9 (1:2:0 - 5183)
No official documentation
Parameters set based on VP9 developers recommendations
Intra Period = 1s
– 24 @ 24 fps, 32 @ 30 fps
19/8/2014
Applications of Digital Image Processing XXXVII
SPIE Optical Engineering + Applications, San Diego, CA, USA
7
8. Encoding
5 different bit rates for each content and codec
– Different spatio-temporal characteristics of the contents
– Both natural and synthetic content
Targeted a fixed QPs mode or equivalent(?)
– Upper bit rate limit: 20 Mbps
Bit rates selected based on expert viewing sessions
19/8/2014
Applications of Digital Image Processing XXXVII
SPIE Optical Engineering + Applications, San Diego, CA, USA
8
9. Methodology
Double Stimulus Impairment Scale (DSIS) Variant II
19/8/2014
Applications of Digital Image Processing XXXVII
SPIE Optical Engineering + Applications, San Diego, CA, USA
9
Age
D
1
Very
annoying
ImperceptibleSlightly
annoying
Perceptible
but not
annoying
Annoying
100
90
80
70
60
50
40
30
20
10
0
10 s
Reference
Video
Test
Video
Reference
Video
Test
Video
10 s 10 s 10 s 5 s2 s 2 s 2 s 2 s = 53s
10. Sessions
ITU BT.500-13
– One test session should not last more than 30 minutes
– Alternate as many different contents as possible
10s sequences => evaluation task requires a lot of attention
– Test sessions no longer than 24 minutes each
– Each session is followed by a resting phase
Details
– Never the same content in consecutive presentations
– Randomization to avoid possible effect of content presentation order
– Dummy sequences at beginning of 1st session to stabilize observers’ rating
– Reference versus reference stimulus to check subject’s reliability
19/8/2014
Applications of Digital Image Processing XXXVII
SPIE Optical Engineering + Applications, San Diego, CA, USA
10
11. Tests
Timing
– 60 sequences (3 codecs × 4 contents × 5 bit rates)
– 3 test sessions + training session
– 1st test session: 2 dummies + 1 ref vs ref + 20 stimuli = 18 x 53 s ~ 21 min
– 2nd test session: 15 stimuli = 20 x 53 s ~ 18 min
– 3rd test session: 15 stimuli = 20 x 53 s ~ 18 min
Test planning
– 1 day, 4 time slots per day
– 6 subjects per slot, split in 2 groups of 3 subjects each
Subjects
– 24 naive subjects, 3F/21M
– Ages ranged from 19 to 35 years old (median of 25.5 years old)
Screening
– Snellen + Ishiara charts
19/8/2014
Applications of Digital Image Processing XXXVII
SPIE Optical Engineering + Applications, San Diego, CA, USA
11
12. Laboratory for subjective video quality assessment
PC server
– SSD solution read and play in real time 3840x2160@30fps
YUV 4:2:0 raw video (i.e., 373.25 MB/s!)
56-inch professional high-performance 4K/QFHD LCD
reference monitor Sony Trimaster SRM-L560
ITU-R BT.500-13 compliant test environment
ITU-R BT.2022 viewing condition
– 3 subjects in front of the screen
– Viewing distance ≈ 1.6x screen height
19/8/2014
Applications of Digital Image Processing XXXVII
SPIE Optical Engineering + Applications, San Diego, CA, USA
12
13. Manege
HEVC vs AVC bit-rate reduction
– BD-PSNR: 28.7%
– BD-MOS: 44.6%
VP9 vs AVC bit-rate reduction
– BD-PSNR: -10.6%
– BD-MOS: -29.2%
HEVC vs VP9 bit-rate reduction
– BD-PSNR: 39.7%
– BD-MOS: 63.7%
19/8/2014
Applications of Digital Image Processing XXXVII
SPIE Optical Engineering + Applications, San Diego, CA, USA
13
14. Traffic
HEVC vs AVC bit-rate reduction
– BD-PSNR: 37.2%
– BD-MOS: 57.5%
VP9 vs AVC bit-rate reduction
– BD-PSNR: -25.1%
– BD-MOS: -61.0%
HEVC vs VP9 bit-rate reduction
– BD-PSNR: 49.8%
– BD-MOS: 74.7%
19/8/2014
Applications of Digital Image Processing XXXVII
SPIE Optical Engineering + Applications, San Diego, CA, USA
14
15. Tree Shade
HEVC vs AVC bit-rate reduction
– BD-PSNR: 22.7%
– BD-MOS: 37.4%
VP9 vs AVC bit-rate reduction
– BD-PSNR: -18.9%
– BD-MOS: 8.2%
HEVC vs VP9 bit-rate reduction
– BD-PSNR: 33.7%
– BD-MOS: 31.9%
19/8/2014
Applications of Digital Image Processing XXXVII
SPIE Optical Engineering + Applications, San Diego, CA, USA
15
16. Sintel2
HEVC vs AVC bit-rate reduction
– BD-PSNR: 69.9%
– BD-MOS: 70.9%
VP9 vs AVC bit-rate reduction
– BD-PSNR: 60.9%
– BD-MOS: 61.7%
HEVC vs VP9 bit-rate reduction
– BD-PSNR: 19.0%
– BD-MOS: 27.5%
19/8/2014
Applications of Digital Image Processing XXXVII
SPIE Optical Engineering + Applications, San Diego, CA, USA
16
17. HEVC vs VP9 vs AVC overall results
General comparison for all content
– average bit-rate reduction over widest range of bit-rates and quality
Robustness in transparent or close to transparent bit-rates
– Further bit-rate savings for each codec based on MOS scores related to fully
transparent or close to transparent quality of decoded content
19/8/2014
Applications of Digital Image Processing XXXVII
SPIE Optical Engineering + Applications, San Diego, CA, USA
17
BD-PSNR BD-MOS BD-PSNR BD-MOS BD-PSNR BD-MOS
39.60% 52.60% 1.59% -5.10% 35.60% 49.40%
HEVC vs AVC VP9 vs AVC HEVC vs VP9
Natural Synthetic Natural Synthetic Natural Synthetic
68.90% 71.90% 43.00% 26.80% 24.30% 56.70%
HEVC AVC VP9
18. Conclusion
Comparison for broadcast scenario – I frame each 1s
Reliability of subjective evaluation results
Variability in codecs performance
– Depending on content and coding conditions
General improvement in compression performance for
HEVC compared to VP9
– When considering wide range of bit-rate form low to high
corresponding to low to transparent quality video content
– Similar performance for higher bit-rates and for synthetic content
Future work
– Comparison between HEVC and VP9 for Internet Streaming
– More content to further verify results
19/8/2014
Applications of Digital Image Processing XXXVII
SPIE Optical Engineering + Applications, San Diego, CA, USA
18
19. 19/8/2014
Applications of Digital Image Processing XXXVII
SPIE Optical Engineering + Applications, San Diego, CA, USA
19
Thank you for your attention!
Hinweis der Redaktion
*Only used for training
Select lower/upper bounds for each content separately
Targeting realistic bit rates
Try to cover the full quality scale
“Rate the level of annoyance of the visual defects that you see in stimulus B, knowing that A is the reference video.”
Training:
Oral instructions to explain the task and the meaning of each label reported on the scale
Viewing session to allow the viewer familiarizing with the assessment procedure
Training samples have quality levels representative of the labels reported on the rating scales