Dual Representations for Light Field Compression

1
Dual Representations for Light Field Compression

• EE368C Project
• January 30, 2001
• Peter Chou
• Prashant Ramanathan

2
Outline

• Background
• Model-based Coding
• Surface Light Fields
• Trade-offs
• Duality
• Proposed Experiments

3
Light Fields and Compression

• What are light fields?
• 2-D array of images
• Why is compression necessary?
• Light fields are very large data sets

Mouse Hemispherical Light Field University of
Erlangen
Michelangelos Night 96 GB raw image
data Stanford Computer Graphics Laboratory
4
Light Fields with Geometry

• Geometry used for light fields to aid compression
• ex. model-based coding
• Light fields are used with geometry for more
  realistic rendering
• ex. surface light fields

5
Model-based Coding

• Model-based Coding of Multi-Viewpoint Imagery
  (Magnor and Girod, VCIP-2000)
• Eigen-Texture Method Appearance Compression
  based on 3D Model (Nishino, Sato, and Ikeuchi,
  CVPR-1999)

http//www.lnt.de/magnor
6
Surface Light Fields

• Surface Light Fields for 3D Photography (Wood et
  al., Siggraph 2000)

http//grail.cs.washington.edu/projects/slf/
7
Surface Light Fields (contd)

• Geometry acquired through range scan
• For each point on surface, a lumisphere
  represents radiance in all directions
• Lumispheres are coded using either
• function quantization (similar to VQ)
• principal function analysis (similar to PCA)

8
Trade-offs

• Textures
• coherency along 4D coordinate directions
• warping introduces artifacts, and possible loss
  of information
• Surface Light Fields
• more intuitive representation for compression
• lumispheres are represented as continuous
  functions

9
Duality

• View-dominant organization (textures)
• Geometry-dominant organization (surface light
  fields)

? Surface Points ?
View 1
View 2
View N
? Views ?
Surface Point 1
Surface Point 2
Surface Point N
10
Proposed Experiments I

• Compare the two organizations for any difference
  in compression results

? Surface Points ?
View 1
View 2
View N
? Views ?
Surface Point 1
Surface Point 2
Surface Point N
11
Proposed Experiments II

• Reparameterize geometry-dominant organization
  using local coordinate system w.r.t. surface
  normals

? Views ?
Surface Point 1
Surface Point 2
Normal Direction View
Surface Point N
12
Proposed Experiments III

• Use image data directly, instead of converting
  from warped texture data

? Views ?
Surface Point 1
Surface Point 2
image pixels
Surface Point N
13
Workplan