MultiLayered Impostors for Accelerated Rendering

1
Multi-Layered ImpostorsforAccelerated Rendering

• Xavier Decoret, iMAGIS
• This is joint work withGernot Schaufler and
  Julie Dorsey at MIT
• and François Sillion at iMAGIS

2
Complex Environments

• Paris
• 411537 vertices
• 137179 triangles
• 32 textures(most 256x256)
• 6.1 MB geometry

3
IBR in RT-Graphics

• Image-based representations simplify the scene
• Images are unaffected by the depicted scenes
  complexity
• Images are fast to render as textured triangles
• Images themselves can be generated with hardware
• This works if a geometric model is available

4
Previous Work

• Pre-generated Representations
• Grossman98, Dally96, Maciel95, Aliaga96, Chen95,
  McMillan95,
• Xiong96, Sillion97, Darsa98, Pulli97, Laveau94,
  Max96,Rafferty98
• Dynamically Updated Representations
• Torborg96, SGI97, Lengyel97, Regan94,Shade96,
  Schaufler96, Mark97, Mann97Aliaga99

5
Impostors (pre-generated)
SGI Performer (Billboards)
Maciel95
Id Software (3D Sprites)
6
Impostors (dynamically generated)
Schaufler95
Schaufler Stürzlinger 96
7
Meshed Impostors Sillion97
Distant Geometry
Depth Map
Discontinuities
Triangulation
8
Artefacts in IBR
Geometry

• Deformation caused by mesh
• Resolution mismatch

9
Artefacts in IBR
Geometry
?

• Incomplete representation
• Rubber sheet effects

10
Artefacts in IBR
Geometry

• Image cracks
• and more
• Static Shading
• no highlights
• no reflections
• No moving Objects

11
Our Contributions
Reduce some of the identified artefacts

• Reducing Rubber Sheet Triangles
• Multi Mesh Impostors
• Reducing distortions to improve quality
• Dynamic Update

12
The database

• The geometry is organized into objects
• We have a set of viewcells

One edge
Another edge
The street graph
13
Model segmentation
14
Single mesh impostor
15
Rubber Sheet Triangles due to parallax
View from above
impostor
viewpoint
16
Locating Rubber Sheet Triangles

• Parallax creates rubber sheets between objects
  when objects overlap in depth

Front view

• In urban walkthrough, parallax is mainly
  horizontal

17
Multi Mesh Impostor

• The critical zone identifies overlaps between 2
  objects

• When overlapping occurs in image space, one
  object can uncover the other one

• If uncovering is too much, objects must not be on
  the same mesh

18
Multi Mesh Impostor
Single mesh
Several meshes

• Quantifying overlapping between 2 objects
• Place those objects in different layers which are
  too distant in depth
• construct a relation graph
• partition the graph

19
Objects and Layers

• Two objects must go into different layers if two
  points on their geometry can be seen under
  sufficiently different viewing angles amin and
  amax.

Object 1
A
Object 2
B
Q
edge
P
20
Relation Graph

• One node per object

• Edges joining overlapping objects
• Coloriate the graph so that joined vertices have
  different colors
• Each color represent a layer
• Non unique

21
Results Example
22
Results Examples
Start point From which the impostor is computed
23
Results Examples
Start point From which the impostor is computed
24
Results Examples
25
Offline vs Online

• Motion hides small artifacts of stored impostors.

• If the user stops, the correct image should
  appear.
• This requires knowledge of the current viewpoint.
• Offline approaches do not have this information

26
Combining preprocessingwith dynamic updates

• Single mesh impostor replace too much geometry to
  be updated.
• Layers and their contents are suitable for
  regeneration of part of the distant model.
• Layers are updated front to backto improve image
  quality
• silhouettes
• distortions
• resolution mismatches

27
System Architecture

• Preprocessing
• Take geometry and view cells
• Find visible geometry for cells
• Split into near and far part
• Create impostors for far part
• Store as scene per view cell
• Walkthrough
• Page in geometry and textures
• Do dynamic updates if possible
• LOD management

28
Results Video (Paris)
Storage requirements 70 Mbs
Computation time 100 edges per hour
Achieved frame rate 50 Hz
29
Future Work

• Smooth transitions between different
  representations
• Automatic generation of street graph from street
  mesh
• Extension of viewcells from edges to areas and
  volumes
• Reverse approach construction of viewcells to
  optimize use of impostors.

30
Acknowledgements

• This is a joint project between iMAGIS, Grenoble
  and MIT, Cambridge.
• It was supported in part by a joint
  collaborative research grant of NSF and INRIA
  (INT-9724005), an Alfred P. Sloan Foundation
  Research Fellowship (BR-3659), and by a grant
  from Intel Corporation.
• The following people contributed to the
  presented results
• Julie Dorsey, François Sillion, Gernot
  Schaufler, Max Chen, Byong Oh Mok, Yann Argotti
  and Sami Shalabi.