Professional Documents
Culture Documents
Marius PREDA
MPEG 3D Graphics Convenor
Institut Polytechnique de Paris, FRANCE
Are we ready for the immersion in digital worlds?
Are we ready for the immersion in digital worlds?
Realistic Computer
Drawings Paintings Paintings Photography
Generated
A more recent history …
New ways
of
consuming
visual
vs
content
Traditional way
of watching
visual content
A more recent history …
New ways
Fixed support Mobile support of
Fixed user position Dynamic user position
consuming
User as a spectator visual
User as a director
vs
Same experience for all content
Personalized experience
Traditional way
of watching
visual content
What is the present and the near future?
The MPEG answer is YES! and even more …, standards are needed!
MPEG started in 2017 MPEG-I (Coded Representation of Immersive Media)
What is the present and the near future?
The MPEG answer is YES! and even more …, standards are needed!
MPEG started in 2017 MPEG-I (Coded Representation of Immersive Media)
Yes, we are!
What are the point clouds and how they enable immersion?
What’s next?
Synergies by the confluence of two worlds of technologies
Visual
capture
Visual
synthesis
Synergies by the confluence of two worlds of technologies
Visual
capture
Multi-camera
LDR, HDR
Stereoscopy
Visual
synthesis
Synergies by the confluence of two worlds of technologies
Visual
capture
Geometric
primitives
Visual
synthesis
Synergies by the confluence of two worlds of technologies
Visual
capture
Multi-camera
LDR, HDR
Geometric
HD, Full HD, 4K, 8K primitives
Stereoscopy
Visual
synthesis
Synergies by the confluence of two worlds of technologies
Visual
capture
Easy to produce
High quality
Interactivity
Immersion
Visual
synthesis
Synergies by the confluence of two worlds of technologies
Visual
capture
Easy to produce
High quality
Interactivity
Immersion
Visual
synthesis
What is at the frontier?
Visual
synthesis
Point Cloud
A set of 3D points
• not ordered,
• without relations between
them
360°
backgroun
d
3D
objects
Realistic content
• spatially collocated,
• interactive
Environment mapping for autonomous driving
~20 million points
• 2,020,734,515 bytes
Point Cloud
0
1 2 3 4
5 6 7 8
9 10 11
12 13 14
Geometry Compressed
Geometry Padded geometry
image geometry video
images
generation images
C om pre sse d
multiple xer
frame
bitstrea m
Smoothed
geometry
Occupancy
Compressed
Occupancy map occupancy
compression map
Compressed
Auxiliary patch-info auxiliary patch
compression information
Video-based Point Cloud Compression
Main ideas:
(1) a point coordinate is encoded as a distance with respect to a
particular plane – inspired from the “displacement mapping” in
Graphics
Main ideas:
(2) the color (or any attribute) associated to a 3D vertex is
encoded in a 2D texture – inspired from the “texture mapping” in
Graphics
Geometry
Color (Attributes)
Video-based Point Cloud Compression
~
Video-based Point Cloud Compression
~
Video-based Point Cloud Compression
MPEG is very
good in video
coding!
Problem solved
Video-based Point Cloud Compression
Frame 24 Frame 26
inter prediction
intra prediction
prediction type
motion vectors
Video-based Point Cloud Compression
inter prediction
intra prediction
prediction type
motion vectors
Video-based Point Cloud Compression
The good thing is that patch forming in out of the scope of the
standard : a good opportunity to have competition at the
encoder level!
Video-based Point Cloud Compression
Up to 20% gain
Video-based Point Cloud Compression
3
10 Decoded
point cloud
Reference B
point cloud
• D1: Point-to-point A
,
, ,
∈ ∈
• D2: Point-to-plane
,
∈ ∈
Decoded
point cloud
10 Reference
point cloud
B
Codec C1
Quality
Codec C2
Bitrate
[1] G. Bjontegaard, Calculation of Average PSNR Differences between RD curves, VCEG-M33, April 2001.
V-PCC Evolution (All Intra): TMC2-v11 vs. TMC2-v1 vs. Anchor[1]
Loot (all intra) Red & black (all intra) Soldier (all intra)
75.0 75.0 75.0
D1-PSNR (dB)
D1-PSNR (dB)
65.0 65.0 65.0
Loot (all intra) Red & black (all intra) Soldier (all intra)
45.0 45.0 44.0
Y-PSNR (dB)
Y-PSNR (dB)
[1] https://github.com/RufaelDev/pcc-
mp3dg/tree/mpeg_standalone_branch
V-PCC Evolution (All Intra): TMC2-v11 vs. TMC2-v1
V-PCC Evolution (Random Access): TMC2-v11 vs. TMC2-v1 vs. Anchor
Loot (random access) Red & black (random access ) Soldier (random access)
75.0 75.0 75.0
D1-PSNR (dB)
D1-PSNR (dB)
D1-PSNR (dB)
Loot (random access) Red & black (random access) Soldier (random access)
45.0 45.0 44.0
Y-PSNR (dB)
Y-PSNR (dB)
Loot (All Intra vs. Random Access) Red & black (All Intra vs. Random Access) Soldier (All Intra vs. Random Access)
73.0 73.0 73.0
D1-PSNR (dB)
D1-PSNR (dB)
D1-PSNR (dB)
69.0 69.0 69.0
Loot (All Intra vs. Random Access) Red & black (All Intra vs. Random Access) Soldier (All Intra vs. Random Access)
45.0 45.0 43.0
Y-PSNR (dB)
Y-PSNR (dB)
35.0 35.0 33.0
Loot (all intra) Red & black (all intra) Soldier (all intra)
74.0 73.0 73.0
D1-PSNR (dB)
D1-PSNR (dB)
70.0 69.0 69.0
Loot (all intra) Red & black (all intra) Soldier (all intra)
45.0 45.0 44.0
39.0
Y-PSNR (dB)
40.0
Y-PSNR (dB)
40.0
Y-PSNR (dB)
Loot (random access) Red & black (random access ) Soldier (random access)
74.0 73.0 72.0
72.0 71.0
D1-PSNR (dB)
D1-PSNR (dB)
D1-PSNR (dB)
70.0
70.0 69.0
TMC2-HEVC 68.0
TMC2-HEVC 67.0
TMC2-HEVC
68.0
Loot (random access) Red & black (random access) Soldier (random access)
45.0 45.0 44.0
40.0 39.0
Y-PSNR (dB)
40.0
Y-PSNR (dB)
Y-PSNR (dB)
Coordinate conversion,
Attribute transfer (for lossy geometry encoding)
Scaling
Geometry
Voxelized source
bitstream
point cloud
(geometry only) Direct coding mode Neighbor-Dependent
Entropy Context
Planar mode
Intra context prediction
QTBT mode
Geometry
Voxelized source
bitstream
point cloud Voxels for the mesh approx. Mesh vertices position
(geometry only)
Trisoup
representation
Branch
vertex with
one child
Leaf vertex
LoD scheme
(Predicting Transform,
Lifting Transform) Quant. coefficient.
Arithmetic coding
Two main methods are provided for the various kind of the attribute data
Geometry-based Point Cloud Compression
Encoding 3D point clouds in their native format
Mesh –
A surface approximation of the point cloud
Visual
synthesis
Future of V-PCC
A set of 3D points
• ordered,
• connected to form
polygons
A set of 3D points
• ordered,
• connected to form
polygons – varying in time
A D-Mesh is defined by
• (Xt, Yt, Zt)n
• (v1t, v2t, v3t)m
• (R, G, B) – still image video
• mapping from video texture to geometry
• reflectance, transparency, …
Automatically captured
V-PCC for the D-Mesh object
Ongoing
exploration in
MPEG
Future of V-PCC
… …
V-PCC
packing/encoding/3D
reconstruction
Future of V-PCC
Classroom sequence
… …
V-PCC
packing/encoding/3D
reconstruction
Conclusions
Novel capturing systems and interactive 3D viewing experiences
are creating new opportunities for realistic immersion
Thank you!