Limn: Using Image and Video Technology for Visualizing a

advertisement
Limn: Using Image and Video
Technology for Visualizing a
Million Cases of Multivariate data
Di Cook, Les Miller,
Manuel Suarez,
Peter Sutherland, Jing Zhang
History
John McDonald’s 3D rotation and linked
brushing in early 90s
Why Scale Visual Methods?
All three data sets have correlation
approximately 0.7
Type of Data
Real-valued multivariate,
spatio-temporal context, class information
Goals
For 100Gb of multivariate data:
Provide multiple scatterplot displays,
either as several static plots, or a
sequence of tour projections.
 Link brush between them in real-time.

Real-time Graphics
GGobi can handle 1 million points, but…
There is a lot of overplotting even when
using pixel sized glyph.
 Brush slow to color with continuous
updating.
 Tour is too slow to tolerate: use
scramble

Problems with Large Data Sets
Data Reduction: MV binning
produces visual artifacts,
screen size induces a
binning, though.
 Scaling of Methods:
computation, storage,
memory.

Screen Resolution
• Limited screen space
induces binning
• Alpha-blending
removes focus from low
density pixels
• Using grey scale
Indexing Projections
nxp data matrix
Pixel resolution plot window
Limn Software
Two Steps:
1.
Create indexing on projected data, or
an animation sequence and save as
QuickTime movie.
2.
View density plots or animation, and
interact with it by brushing, and
overlaying subsets of data.
* Code is written in Java
Data Description
Seasonal metrics data computed on
AVHRR images across the USA in 1989:
SoST, SoSN, EoST,EoSN, MaxT, MaxN,
TotalNDVI.
And land cover classes: agriculture,
grassland, deciduous, evergreen, barren.
Demo
Scatterplot matrix
 Tour movie

Currently we’re also exploring…
Sampling distributions of the pixel
population.
 Overlaying movies for working with
larger data.
 Broader array of data parsers.

Contact Information
Software is archived at:
www.sourceforge.net
Email: dicook@iastate.edu
Download