Page 1
www.compaq.com Mitsubishi ElectricResearch Laboratory
www.merl.com
Feline: Fast, Elliptical Feline: Fast, Elliptical Lines for Anisotropic Lines for Anisotropic Texture MappingTexture Mapping
Joel McCormack, Ronald Perry, Joel McCormack, Ronald Perry, Keith I. Farkas, and Norman P. Keith I. Farkas, and Norman P. JouppiJouppi
Page 2
www.compaq.com Mitsubishi ElectricResearch Laboratory
www.merl.com
OutlineOutlineWhat should texture mapping really do?Previous workOur work: Simple FelineImage galleryConclusions
Page 3
www.compaq.com Mitsubishi ElectricResearch Laboratory
www.merl.com
An Intuitive Diagram of An Intuitive Diagram of Texture Mapping (stolen Texture Mapping (stolen from Landsdale)from Landsdale)
Pixel filter is a window with variable transparency
Filter weight at a point is degree of transparency
A circular window views an elliptical texture area
Adjacent filters overlap
Page 4
www.compaq.com Mitsubishi ElectricResearch Laboratory
www.merl.com
Heckbert & Greene’s Heckbert & Greene’s Elliptical Weighted Average Elliptical Weighted Average (EWA)(EWA)
0.000
0.005
0.010
0.015
0.020
0.025
0.030
0.035
SampleWeight
u axis
v axis0.000
0.005
0.010
0.015
0.020
0.025
0.030
0.035
u axis
v axis
Assume perspective distortion is constant near pixel (99.9% true)
Back mapping to screen space is a biquadraticPerspective Warp Affine Warp
Page 5
www.compaq.com Mitsubishi ElectricResearch Laboratory
www.merl.com
EWA Texturing ExamplesEWA Texturing Examples
Few & faint Moiré artifacts
Nice definition to text without jaggies
Page 6
www.compaq.com Mitsubishi ElectricResearch Laboratory
www.merl.com
Lance Williams’ Mip-Lance Williams’ Mip-mapping with Trilinear mapping with Trilinear FilteringFiltering
Trilinear filter has (very) roughly circular contour lines, with a square footprint
0.000
0.005
0.010
0.015
0.020
0.025
0.030
0.035
u axis
v axis0.000
0.005
0.010
0.015
0.020
0.025
0.030
0.035
SampleWeight
u axis
v axis
Affine Warp Trilinear Filter
Page 7
www.compaq.com Mitsubishi ElectricResearch Laboratory
www.merl.com
Trilinear Texturing Trilinear Texturing ExamplesExamples
Ellipticalfootprint
Trilinearfootprint
Ellipticalfootprint
Trilinearfootprint
Page 8
www.compaq.com Mitsubishi ElectricResearch Laboratory
www.merl.com
0.0000.0050.0100.0150.020
0.025
0.030
0.035
0.040
0.045
u axis
v axis
Texram (Best Case) Texram (Best Case) 2n trilinear probes along ellipse’s “major axis”Resulting filter is too short and mesa-like at best...
0.000
0.005
0.010
0.015
0.020
0.025
0.030
0.035
SampleWeight
u axis
v axis
Affine WarpTexram (best case)
Page 9
www.compaq.com Mitsubishi ElectricResearch Laboratory
www.merl.com
0.0000.0050.0100.0150.020
0.025
0.030
0.035
0.040
0.045
u axis
v axis
Texram (Worst Case) Texram (Worst Case) … and jaggedly peaked at worst.
0.000
0.005
0.010
0.015
0.020
0.025
0.030
0.035
SampleWeight
u axis
v axis
Affine WarpTexram (worst case)
Page 10
www.compaq.com Mitsubishi ElectricResearch Laboratory
www.merl.com
Texram Texturing ExamplesTexram Texturing Examples
Lots of swimming Moiré artifacts
Sharp text but with jaggies (aliasing)
Page 11
www.compaq.com Mitsubishi ElectricResearch Laboratory
www.merl.com
Exact Feline: Use Ellipse Exact Feline: Use Ellipse ParametersParameters
n probes, Gaussian weighted, on major axis of ellipseImportant Texram problems fixedBut expensive setup comparable to EWA!
0.000
0.005
0.010
0.015
0.020
0.025
0.030
0.035
SampleWeight
u axis
v axis
Affine Warp
0.000
0.005
0.010
0.015
0.020
0.025
0.030
0.035
Exact Feline
Page 12
www.compaq.com Mitsubishi ElectricResearch Laboratory
www.merl.com
Simple Feline: Approximate Simple Feline: Approximate EllipseEllipse
Approximate ellipse axes for cheaper setupSlightly underestimates major radius, overestimates minor radius (resulting in fewer, fatter probes)
0.000
0.005
0.010
0.015
0.020
0.025
0.030
0.035
SampleWeight
u axis
v axis0.000
0.005
0.010
0.015
0.020
0.025
0.030
0.035
u axis
v axis
Affine Warp Simple Feline
Page 13
www.compaq.com Mitsubishi ElectricResearch Laboratory
www.merl.com
Simple Feline: Reducing # Simple Feline: Reducing # of Probesof Probes
Allow shortening of “major axis”Extreme sensitivity, quickly causes aliasing3% decrease
Allow widening probesCauses blurringMax of 16% (high-quality)or 31% (high-efficiency)
Allow stretching distance between probesCauses aliasingMax of 15% (high-quality) or 36% (high-efficiency)
Page 14
www.compaq.com Mitsubishi ElectricResearch Laboratory
www.merl.com
Simple Feline: Modest Simple Feline: Modest Probe ReductionProbe Reduction
Actual blur stretch of 15.6% (maximum allowed)Actual aliasing stretch of 7.0%
0.000
0.005
0.010
0.015
0.020
0.025
0.030
0.035
SampleWeight
u axis
v axis0.000
0.005
0.010
0.015
0.020
0.025
0.030
0.035
u axis
v axis
Affine Warp High-Quality Simple Feline
Page 15
www.compaq.com Mitsubishi ElectricResearch Laboratory
www.merl.com
0.000
0.005
0.010
0.015
0.020
0.025
0.030
0.035
u axis
v axis
Simple Feline: “Texram Simple Feline: “Texram Probe-Equivalent” Probe-Equivalent” ReductionReduction
Actual blur stretch of 31% (maximum allowed)Actual aliasing stretch of 26%
0.000
0.005
0.010
0.015
0.020
0.025
0.030
0.035
SampleWeight
u axis
v axis
Affine Warp High-Efficiency Simple Feline
Page 16
www.compaq.com Mitsubishi ElectricResearch Laboratory
www.merl.com
0.000
0.005
0.010
0.015
0.020
0.025
0.030
0.035
u axis
v axis
Simple Feline: “Texram Simple Feline: “Texram Probe-Equivalent” with Probe-Equivalent” with Gaussian ProbesGaussian Probes
Blur stretch of 36%Aliasing stretch of 31%
0.000
0.005
0.010
0.015
0.020
0.025
0.030
0.035
SampleWeight
u axis
v axis
Affine Warp High-Efficiency Feline w/Gaussian
Slightly blurrier, but many fewer aliasing artifacts
Page 17
www.compaq.com Mitsubishi ElectricResearch Laboratory
www.merl.com
Image Gallery: TextImage Gallery: Text
HQ Feline EWA
Trilinear Texram
Page 18
www.compaq.com Mitsubishi ElectricResearch Laboratory
www.merl.com
Image Gallery: CurvesImage Gallery: Curves
EWAHQ Feline
Trilinear Texram
Page 19
www.compaq.com Mitsubishi ElectricResearch Laboratory
www.merl.com
Image Gallery: Image Gallery: CheckerboardsCheckerboards
HQ Feline EWA
Trilinear Texram
Page 20
www.compaq.com Mitsubishi ElectricResearch Laboratory
www.merl.com
ConclusionsConclusionsFeline compared to EWA:
High-Quality Feline’s visual quality is comparable, using half the cycles/pixel and much smaller setup logic
Feline compared to Texram:High-quality images far superior, using more cycles/pixel
High-efficiency images superior, using same cycles/pixel
Feline requires a tiny fraction of the die of a PC or next-generation game console graphics accelerator