Managing Shared Virtual Worlds: Here, There, Everywhere

1
Managing Shared Virtual Worlds Here, There,
Everywhere

• Michael V. Capps
• MIT Laboratory for Computer Science
• Mitsubishi Electric Research Laboratory

2
Overview
Motivation
Questions?
Here and now
Here Tomorrow
Conclusions
Everywhere
Building a Model
There
3
VR not enough R

• low-resolution, low-FOV, cartoon worlds
• displaying
• too much
• too little
• models outpacing technology
• Bad VR is a research area!

Does it look real to you?
4
Networked VR

• Distributed model and client separated
• Collaborative interaction among users
• bandwidth limitation
• poor use, preloading
• VRML Quake

MSIE / VRML
ATT Ring System
5
Solutions?

• no easy solution coming anytime soon
• for now, optimize intelligently
• Goal smooth degradation
• managing whats here what to display
• getting it here from there what to fetch
• interaction everywhere what to share

6
Here--and Now

• goals photo-realism, presence
• most solutions require extremely complex geometry
• and high-fidelity lighting (radiosity / ray
  tracing)
• reducing the work
• visibility
• texturing
• levels of detail
• image-based rendering

7
Here Tomorrow making here a little tougher

• management well explored in limited instances
• LOD in height fields
• LOD in 2 1/2 D axial architectural models
• managing local model is getting tougher
• more general worlds
• hybrid rendering schemes

8
Hybrid Systems

• the best of both worlds?
• use a combination of geometry, textures,
  images...
• examples
• CityScape Xiong
• color-cubes U. Wash
• shared-frustum stereo
• hierarchical image caching
• image impostors

Color-cubes
9
Hierarchical Image Caching

• based on explorations by U. Wash. (SIGGRAPH 96)
• BSP nodes replaced with cached images
• safety zone of angular disparity

Above geometry only Below w/cached images
10
Image Impostors

• Ville visualization of urban scenery
• large model extent
• exploit structured nature
• replaces far-field geometry with impostor mesh
• S. Shalabi, J. Dorsey MIT
• F. Sillion, IMAGIS

11
AnySystem

• Flexible architecture for investigating hybrid
  system design and optimization
• AnyRep
• AnyRend
• AnyTree

Rep
Geom
Image
RepList
Tris
Depth
Impostor
Hi-order
Volume
LightField
Cache
TriList
OptTris
Ht Field
12
AnyTree

• combination of spatial subdivisions
• each optimal in certain situations
• greater generality
• split is very complex

OCT Tree
QUAD Tree
BSP Tree
KD Tree
13
Building a Model

• cant solve can optimize
• need a model that takes into account
• whats most important
• capabilities of the client
• user interest
• perceptual difference between alternatives
• continuous, or at least highly-sampled discrete,
  space of possibilities

14
QuICk Model
Quality of representations u Importance of
representations/areas Cost of representations k
15
Q is for Quality

• rep quality is mostly precomputed / predetermined
• comparing apples and oranges
• essentially discrete quantification of perceptual
  difference
• function of client display technology!
• Current methods
• number of triangles
• find difference by sampling points
• is dynamic based on interfaces with other reps
• in space (stitching)
• in time (hysteresis)

16
Importance

• again mostly a precomputation
• viewpoint-dependent
• designer input
• small addition to model creation time
• plain old distance

17
Other contributors to I

• visitation frequency
• visibility
• point-to-point visibility
• cell-to-cell visibility
• PVS (lower storage)
• GVS (continuous data)
• morphology
• urban structure (Ville)
• motion prediction

18
Cost

• render cost
• based on client capability
• which is determined empirically
• which can be dynamic
• and complexity of rep
• number of polygons, size of image, etc.
• storage cost
• cache management (model gtgt memory)
• disk transfer path latency requires prediction
• Berkeley Walkthrough
• creation cost

19
Using the QuICk model

• Factors are complex, but easy to estimate
• model is a function of client, so
• important to have accurate client
  specification
• MAX(Q?I) so that each area has a rep
  while SUM(C) lt display limit

20
There

• problem reliance on network
• bandwidth limitations
• must choose which areas of model to download
• must choose which representations of those areas
  to download
• latency limitations
• must make predictions to request in advance
• ...sound familiar?

21
QuICk model

• Distribution is complex caching
• just another transfer path
• server optimizes globally
• new QuICk parameters

• Cost
• include additional latency cost of fetch
• rep creation at server

• Importance
• frequency of request
• arrival time vs. time needed
• prediction
• classes of motion
• precomputed distances
• certainty of predictions

22
Everywhere Shared VR systems

• Interest Management today
• distance acceptable standard
• cells
• for distance
• for visibility
• locales
• auras
• functional participant groups
• implicit by client specification
• explicit user-administered

23
Interest by Communication

• position updates
• ephemeral actions
• visible
• audible
• world-state modifications

24
QuICk model

• QuICk still applies (with a stretch)
• Quality accuracy of perception
• of maximum update rate
• per action
• per participant interface
• Importance ? interest
• Cost
• cost of transmission
• cost to process and display

25
Conclusion and Future Work

• Here
• systems have a long way to go
• previous work only scratched surface of system
  management issues
• There
• management is extension of single-user problem
• Everywhere
• interest management has been successful...
• ...but if the model has been built, use it!

26
Conclusion and Future Work

• weve built a generic toolkit to explore
  management of hybrid VR systems
• and have expanded that to multi-user hybrid VR
  systems

27
Questions?

• More information at
• 
  http//gfx.lcs.mit.edu/capps
• Acknowledgements
• -Funded in part by a National Science Foundation
  Graduate Fellowship
• -Thanks for images Sami Shalabi, University of
  Washington, Brian Ladd, IdSoftware, ATT