antialiasing with line samples thouis r. jones, ronald n. perry merl - mitsubishi electric research...
TRANSCRIPT
![Page 1: Antialiasing with Line Samples Thouis R. Jones, Ronald N. Perry MERL - Mitsubishi Electric Research Laboratory](https://reader035.vdocument.in/reader035/viewer/2022070407/56649e175503460f94b02fa4/html5/thumbnails/1.jpg)
Antialiasing with Line Samples
Thouis R. Jones, Ronald N. Perry
MERL - Mitsubishi Electric Research Laboratory
![Page 2: Antialiasing with Line Samples Thouis R. Jones, Ronald N. Perry MERL - Mitsubishi Electric Research Laboratory](https://reader035.vdocument.in/reader035/viewer/2022070407/56649e175503460f94b02fa4/html5/thumbnails/2.jpg)
Antialiasing
• Fundamentally a sampled convolution:
Filter ngBandlimiti ),(
Data Continuous),(
y)at x, (pixel Image Sampled),(
),(),(),(
yxF
yxI
PPS
dxdyPyPxFyxIPPS
yx
yxyx
![Page 3: Antialiasing with Line Samples Thouis R. Jones, Ronald N. Perry MERL - Mitsubishi Electric Research Laboratory](https://reader035.vdocument.in/reader035/viewer/2022070407/56649e175503460f94b02fa4/html5/thumbnails/3.jpg)
Analytic Antialiasing
• Analytic antialiasing requires solving visibility to give a continuous 2D image– Visible polygons tessellate image plane – Arbitrarily complex shapes– Efficient methods exist for evaluating the
integral from 2D tessellation (Duff 1989, McCool 1995)
![Page 4: Antialiasing with Line Samples Thouis R. Jones, Ronald N. Perry MERL - Mitsubishi Electric Research Laboratory](https://reader035.vdocument.in/reader035/viewer/2022070407/56649e175503460f94b02fa4/html5/thumbnails/4.jpg)
Reduce Dimensionality - Point Sampling
• Point sampling reduces the dimension of the visibility calculation to 0D in image plane
• Pixel’s value is a weighted sum of values at sample points in the image plane
• Most widespread and well studied method for antialiasing geometry
![Page 5: Antialiasing with Line Samples Thouis R. Jones, Ronald N. Perry MERL - Mitsubishi Electric Research Laboratory](https://reader035.vdocument.in/reader035/viewer/2022070407/56649e175503460f94b02fa4/html5/thumbnails/5.jpg)
Reduce Dimensionality -1D Sampling
• Another option is to reduce the dimension by 1, and sample along 1D elements
• Prior Art:– Max 1990 - Antialiasing Scan-Line Data– Guenter & Tumblin 1996 - Quadrature
Prefiltering for High Quality Antialiasing– Tanaka & Takahashi 1990 - Cross Scanline
Algorithm
![Page 6: Antialiasing with Line Samples Thouis R. Jones, Ronald N. Perry MERL - Mitsubishi Electric Research Laboratory](https://reader035.vdocument.in/reader035/viewer/2022070407/56649e175503460f94b02fa4/html5/thumbnails/6.jpg)
Prior Art - Max
• Sample along scanlines
• Analytic antialiasing in scanline direction, supersampling in other direction
• Extended in same paper to use edge slopes to better approximate 2D image before 2D filtering
![Page 7: Antialiasing with Line Samples Thouis R. Jones, Ronald N. Perry MERL - Mitsubishi Electric Research Laboratory](https://reader035.vdocument.in/reader035/viewer/2022070407/56649e175503460f94b02fa4/html5/thumbnails/7.jpg)
Prior Art - Guenter & Tumblin
• Quadrature prefiltering - accurate numerical approximation of antialiasing integral
• Assumes existing 2D visibility solution– Phrased as an efficient computation of the
antialiasing integral, not as a sampling method– As in Max 1990, unidirectional sampling
![Page 8: Antialiasing with Line Samples Thouis R. Jones, Ronald N. Perry MERL - Mitsubishi Electric Research Laboratory](https://reader035.vdocument.in/reader035/viewer/2022070407/56649e175503460f94b02fa4/html5/thumbnails/8.jpg)
Prior Art - Tanaka & Takahashi
• Uses horizontal scanlines and vertical “sub-scanlines” to find 2D visibility solution
• Filters 2D image
• Again, not really phrased as a sampling method
![Page 9: Antialiasing with Line Samples Thouis R. Jones, Ronald N. Perry MERL - Mitsubishi Electric Research Laboratory](https://reader035.vdocument.in/reader035/viewer/2022070407/56649e175503460f94b02fa4/html5/thumbnails/9.jpg)
Line Sampling
• Small 1D samples - “line samples”– Centered at pixel, spanning filter footprint
• Multiple line samples and sampling directions per pixel
![Page 10: Antialiasing with Line Samples Thouis R. Jones, Ronald N. Perry MERL - Mitsubishi Electric Research Laboratory](https://reader035.vdocument.in/reader035/viewer/2022070407/56649e175503460f94b02fa4/html5/thumbnails/10.jpg)
Line Sampling (continued)
• 1D filtering only - Cheaper/Faster– 1D tables– Edge slopes ignored in filtering
• Blending of samples based on image features– Does use edge slopes...– …but separates blending from filtering, keeping
both simple
![Page 11: Antialiasing with Line Samples Thouis R. Jones, Ronald N. Perry MERL - Mitsubishi Electric Research Laboratory](https://reader035.vdocument.in/reader035/viewer/2022070407/56649e175503460f94b02fa4/html5/thumbnails/11.jpg)
Theory and Practice
• Theory: – Arbitrary number of line samples per pixel in
arbitrary directions
• Practice:– 2 line samples per pixel, horizontal and vertical– Line samples are subsegments of horizontal and
vertical “scanlines”
![Page 12: Antialiasing with Line Samples Thouis R. Jones, Ronald N. Perry MERL - Mitsubishi Electric Research Laboratory](https://reader035.vdocument.in/reader035/viewer/2022070407/56649e175503460f94b02fa4/html5/thumbnails/12.jpg)
Practice (continued)
![Page 13: Antialiasing with Line Samples Thouis R. Jones, Ronald N. Perry MERL - Mitsubishi Electric Research Laboratory](https://reader035.vdocument.in/reader035/viewer/2022070407/56649e175503460f94b02fa4/html5/thumbnails/13.jpg)
Line Sampling Algorithm
• Determine visible segments along line samples at each pixel
• Keep sum of weights at each pixel (from edge crossings)
• Apply 1D table-based filter
• Blend values from vertical and horizontal line samples
![Page 14: Antialiasing with Line Samples Thouis R. Jones, Ronald N. Perry MERL - Mitsubishi Electric Research Laboratory](https://reader035.vdocument.in/reader035/viewer/2022070407/56649e175503460f94b02fa4/html5/thumbnails/14.jpg)
Determining Visible Segments
• Horizontal and vertical line samples are subsegments of scanlines– Use scanline methods for visibility
• Less efficient methods for arbitrary sampling directions (see paper)
![Page 15: Antialiasing with Line Samples Thouis R. Jones, Ronald N. Perry MERL - Mitsubishi Electric Research Laboratory](https://reader035.vdocument.in/reader035/viewer/2022070407/56649e175503460f94b02fa4/html5/thumbnails/15.jpg)
Weights from Edge Crossings
• Why edge crossings?
– A line sample’s accuracy depends on its orientation relative to image features
– If a line sample intersects an edge, its filtering accuracy is highest when perpendicular, lowest when parallel
![Page 16: Antialiasing with Line Samples Thouis R. Jones, Ronald N. Perry MERL - Mitsubishi Electric Research Laboratory](https://reader035.vdocument.in/reader035/viewer/2022070407/56649e175503460f94b02fa4/html5/thumbnails/16.jpg)
Weights (continued)
• Use as weight– Normalized: weights for
horizontal and verticalline samples sum to one
2sin
![Page 17: Antialiasing with Line Samples Thouis R. Jones, Ronald N. Perry MERL - Mitsubishi Electric Research Laboratory](https://reader035.vdocument.in/reader035/viewer/2022070407/56649e175503460f94b02fa4/html5/thumbnails/17.jpg)
Weights (continued)
• Sum weights at each pixel (post-visibility)
• Intersecting triangles - use cross product of normals to find slope of created edge
• Edge weights should be adjusted by color change across the edge
![Page 18: Antialiasing with Line Samples Thouis R. Jones, Ronald N. Perry MERL - Mitsubishi Electric Research Laboratory](https://reader035.vdocument.in/reader035/viewer/2022070407/56649e175503460f94b02fa4/html5/thumbnails/18.jpg)
1D Table-based Filter
• Stretch 1D to 2D, then filter• Perpendicular, not according to edge slope
• Combine stretch and filter– Use summed filter table
![Page 19: Antialiasing with Line Samples Thouis R. Jones, Ronald N. Perry MERL - Mitsubishi Electric Research Laboratory](https://reader035.vdocument.in/reader035/viewer/2022070407/56649e175503460f94b02fa4/html5/thumbnails/19.jpg)
Blend Values from Line Samples
• Sample weights are
• Good results using step function for blending, but discontinuity can cause aliasing
• Use cubic blending (Hermite)
2sin
![Page 20: Antialiasing with Line Samples Thouis R. Jones, Ronald N. Perry MERL - Mitsubishi Electric Research Laboratory](https://reader035.vdocument.in/reader035/viewer/2022070407/56649e175503460f94b02fa4/html5/thumbnails/20.jpg)
Results
![Page 21: Antialiasing with Line Samples Thouis R. Jones, Ronald N. Perry MERL - Mitsubishi Electric Research Laboratory](https://reader035.vdocument.in/reader035/viewer/2022070407/56649e175503460f94b02fa4/html5/thumbnails/21.jpg)
Horizontal Filtering Only
![Page 22: Antialiasing with Line Samples Thouis R. Jones, Ronald N. Perry MERL - Mitsubishi Electric Research Laboratory](https://reader035.vdocument.in/reader035/viewer/2022070407/56649e175503460f94b02fa4/html5/thumbnails/22.jpg)
Comparison - Radial Triangles16x Supersampling
![Page 23: Antialiasing with Line Samples Thouis R. Jones, Ronald N. Perry MERL - Mitsubishi Electric Research Laboratory](https://reader035.vdocument.in/reader035/viewer/2022070407/56649e175503460f94b02fa4/html5/thumbnails/23.jpg)
Comparison - Radial Triangles256x Supersampling
![Page 24: Antialiasing with Line Samples Thouis R. Jones, Ronald N. Perry MERL - Mitsubishi Electric Research Laboratory](https://reader035.vdocument.in/reader035/viewer/2022070407/56649e175503460f94b02fa4/html5/thumbnails/24.jpg)
Comparison - Triangle Comb16x 256x Line Sampling
![Page 25: Antialiasing with Line Samples Thouis R. Jones, Ronald N. Perry MERL - Mitsubishi Electric Research Laboratory](https://reader035.vdocument.in/reader035/viewer/2022070407/56649e175503460f94b02fa4/html5/thumbnails/25.jpg)
Comparison - Animation
![Page 26: Antialiasing with Line Samples Thouis R. Jones, Ronald N. Perry MERL - Mitsubishi Electric Research Laboratory](https://reader035.vdocument.in/reader035/viewer/2022070407/56649e175503460f94b02fa4/html5/thumbnails/26.jpg)
Benefits of Line Sampling
• High quality
• Near analytic for substantially vertical or horizontal edges
• Low variance near lone edges
• Efficient– 2 scanline passes + 1D filtering + blending
![Page 27: Antialiasing with Line Samples Thouis R. Jones, Ronald N. Perry MERL - Mitsubishi Electric Research Laboratory](https://reader035.vdocument.in/reader035/viewer/2022070407/56649e175503460f94b02fa4/html5/thumbnails/27.jpg)
Failure Cases
• Areas with high frequency content in two directions
– Small features can be missed
– Corners
• Non-trivial to extend to curved surfaces
![Page 28: Antialiasing with Line Samples Thouis R. Jones, Ronald N. Perry MERL - Mitsubishi Electric Research Laboratory](https://reader035.vdocument.in/reader035/viewer/2022070407/56649e175503460f94b02fa4/html5/thumbnails/28.jpg)
Conclusions and Future Work
• Line Sampling can provide near-analytic quality antialiasing at substantially lower cost
• Future work:– Implement in realtime scanline renderer– Integration with texture mapping– Stochastic line sampling– Extension to motion blur– Reduced memory requirements
![Page 29: Antialiasing with Line Samples Thouis R. Jones, Ronald N. Perry MERL - Mitsubishi Electric Research Laboratory](https://reader035.vdocument.in/reader035/viewer/2022070407/56649e175503460f94b02fa4/html5/thumbnails/29.jpg)
Acknowledgements
• Nelson Max
• Rob Kotredes
• Richard Coffey
• David Hart
• Peter-Pike Sloan
• MERL: Hanspeter Pfister, Larry Seiler, Joe Marks