Image-Based Rendering CS 446: Real-Time Rendering

1
Image-Based Rendering CS 446 Real-Time
Rendering Game Technology

• David Luebke
• University of Virginia

2
Demo

• Today John Dimeo
• Thursday Meng Tan

3
Image-Based Rendering

• Youve been learning how to turn geometric models
  into images
• Specifically, images of compelling 3D objects and
  worlds
• Image-based rendering a relatively new field of
  computer graphics devoted to making images from
  images
• Ex Quicktime VR

4

• Quicktime VR

5
Images with depth

• Quicktime VR is really just a 2D panoramic
  photograph
• Spin around, zoom in and out
• But what if we could assign depth to parts of the
  image?
• Ex Tour Into the Picture

6

• Tour Into the Picture

7
Tour Into the Picture

• Software for
• Selecting parts of an image
• Assigning a vanishing point for depth of
  background objects
• Assigning depth to foreground objects
• Painting in behind objects

8
Image-Based Modeling and Editing Byong Mok Oh,
Max Chen, Julie Dorsey, and Fredo Durand
9
Depth per pixel

• What if we could assign an exact depth to every
  pixel?
• Ex MIT Image-Based Editing system

10
Depth per pixel continued

• What if we had a camera that automatically
  acquired depth at every pixel?
• Ex deltasphere
• Ex Monticello project

11
Laser rangefinder scanner

• Deltasphere 3000 by 3rdTech
• Time of flight laser rangefinder
• Infrared or red visible
• 20,000 samples per second
• 40 foot range
• Accuracy 1 mm
• Panoramic scanner with spinning mirror to scan
  all directions
• High-resolution digital camera with same nodal
  point
• Many similar products for different niches

12
An aside Virtual Monticello

• Switch presentations

13
The Goal From this
14
to this
15
to this
16
to this.
17
Scanning MonticelloChallenges

• Single scan 5-10 million points
• Each room 4-6 scans
• Jeffersons private suite 5 rooms

18
Point cloud ? mesh

• Simplest approach connect the dots
• (Demo)

19
Point cloud ? mesh

• Problems
• Shouldnt always fill holes
• Need to merge multiple scans

20
Distance field

• Instead build volumetric distance field, extract
  zero-valued isosurface
• Robust to small errors within and between scans
• Guaranteed to produce watertight mesh

21
Distance fieldsEfficient construction

• Usually represent DF on uniform grid
• Naive approach to constructing DF
• Visit every grid cell (voxel)
• Find nearest polygon in mesh
• Record distance to voxel center
• So whats the problem?
• Would like at least 1024x1024x1024 voxels
• 5-10 million polygons in mesh

22
Distance fieldsEfficient construction

• Could instead represent DF hierarchically
• Our approach
• Walk through the scan file, inserting points into
  grid
• Calculate local normal vector to mesh
  incrementally
• Propagate distance from mesh outward from voxel
  containing point, to some maximum ramp width
• Use low-res 1-bit occupancygrid to indicate
  which voxelsintersect mesh
• Walk through occupied voxelsto produce mesh

23
Some Results
A single scan of Monticello library room (after
range noise reduction)
24
Some Results
Reconstructed model 2.86 million vertices, 5.53
million triangles. Simplified to 1 million
triangles.
25
Some Results