 |
|
|
Feline: Fast Elliptical Lines for Anisotropic Texture
Mapping |
|
|
Joel McCormack*, Ronald Perry**, Keith I. Farkas*, and
Norman P. Jouppi*
|
|
|
*Compaq Computer Corporation's Western Research Lab
|
**Mitsubishi Electric Research Laboratory
|
|
{Joel.McCormack, Keith.Farkas, Norm.Jouppi}@compaq.com
|
perry@merl.com
|
This web page describes in greater detail the parameters used to generate the images that appeared in the proceedings of SIGGRAPH 99 and provides links to the actual images. This page also includes links to some images that do not appear in the paper but which provide some additional insight into Feline and how it compares to other algorithms. Please note that these images should be viewed on a gamma corrected display, or using software that will gamma correct the images before displaying them.
Further information about Feline and other related algorithms can be obtained from the technical report entitled "Simple and Table Feline: Fast Elliptical Lines for Anisotropic Texture Mapping", which is available from WRL's web site.
(Not shown in the paper.) Simple Feline using the constant rounding of iProbes described in Section 3.1, and no fudge factors, using a trilinear probe. Mip-maps were constructed with a box filter. Aliasing artifacts abound, as the base texture and all mip-maps contain frequencies that are too high for the narrow probe filter to remove.
(Not shown in the paper.) Simple Feline using the constant rounding of iProbes described in Section 3.1, and no fudge factors, using a trilinear probe. Mip-maps were constructed with a Lanczos filter of radius 3. Aliasing artifacts are substantially reduced, as the mip-maps now contain more reasonable frequencies.
(Not shown in the paper.) Mip-mapped EWA (with a radius between 2 and 4 texels, which is about twice as slow as the radius 1.5 to 3 algorithm used elsewhere.) Mip-maps were constructed with a box filter. Due to the relatively wide filter, aliasing artifacts are much less strong than the equivalent Feline image, though still quite visible.
(Not shown in the paper.) Mip-mapped EWA (with a radius between 2 and 4 texels). Mip-maps were constructed with a radius 3 Lanczos filter. The Lanczos further reduces some of the aliasing, but we found it impossible to use: in other images the Lanczos filter makes EWA's transition from a large ellipse on one mip-map level to a small ellipse (1/2 the size in each dimension) on the next mip-map level stunningly visible.
(Figure 10 in the paper.) Trilinear filtering paints curved lines with a significant amount of blurring. This image was generated by a standard trilinear filter, with perfect level-of-detail computation using the maximum of the two vectors.
(Figure 11 in the paper.) Texram paints curved lines with the strong Moiré artifacts.
(Not shown in the paper.) High-efficiency Simple Feline using a trilinear probe filter paints the curved lines with fewer artifacts than Texram.
(Figure 12 in the paper.) High-efficiency Simple Feline using a Gaussian probe filter paints the curved lines with fewer artifacts than with a trilinear probe filter. Note in particular the disappearance of the very faint trilinear probe artifacts in the lower right corner of the upper left square in the previous figure.
(Not shown in the paper.) High-quality Simple Feline using a trilinear probe paints the curved lines with few artifacts.
(Figure 13 in the paper.) High-quality Simple Feline using a Gaussian probe paints the curved lines with few artifacts. In this case, the Gaussian probe causes slightly more blurriness than the trilinear probe.
(Not shown in the paper.) Simple Feline using the constant rounding of iProbes described in Section 3.1, and no fudge factors, using a trilinear probe. We could discern no difference in image quality compared to high-quality Feline with a trilinear probe, but this image required 31% more probes.
(Not shown in the paper.) Simple Feline using the constant rounding of iProbes and a Gaussian probe. Again, the Gaussian creates a slightly blurrier image.
(Figure 15 in the paper.) Texram paints bricks with herringbone artifacts. Especially when animated, the area in the herringbone pattern appears to breathe: the area seems to rise up on a small hill, which then sinks back into the level plane.
(Not shown in the paper.) High-efficiency Simple Feline with a trilinear probe paints bricks with fewer artifacts, though a faint herringbone pattern remains.
(Figure 16 in the paper.) High-efficiency Simple Feline with a Gaussian probe paints bricks with fewer artifacts. The Gaussian probe reduces the herringbone pattern to near invisibility.
(Not shown in the paper) High-quality Simple Feline with a trilinear probe paints bricks with few artifacts, but note the small faint herringbone pattern near the horizon.
(Figure 17 in the paper) High-quality Simple Feline with a Gaussian probe paints bricks with few artifacts. The Gaussian almost eliminates the faint herringbone pattern visible in the previous image. This presented us with a small dilemma: the high-quality Feline curved lines look sharper using a trilinear probe, but the bricks exhibit fewer aliasing artifacts using a Gaussian probe. We chose to use the Gaussian probe for all Feline images in the paper.
(Not shown in the paper) Constant-rounding Simple Feline with a trilinear probe paints bricks with few artifacts, and is sharper than the high-quality Gaussian probe. Again, constant-rounding requires 31% more probes than high-quality Feline. Is the sharpness worth the extra cycles?
(Not shown in the paper)) Constant-rounding Simple Feline with a Gaussian probe simply blurs the previous image.
(Figure 19 in the paper.) Trilinear filtering paints blurry text.
(Not shown in the paper.) Trilinear filtering sharpness depends upon the level-of-detail computed, which determines the radius of the filter. Usually, the maximum length of the two vectors are used. In this image, we used the minimum length of the two vectors. It's sharper all right :)
((Figure 20 in the paper.) Texram paints text with some stairstepping. In the left column of letters, note the roughness of the digits and the "@" sign. Further up, characters become outright distorted; see the "C" for example.
(Figure 21 in the paper.) High-efficiency Simple Feline with a Gaussian probe paints smooth text. In the Feline series of text images, no aliasing artifacts are visible. To discern any differences, get back from the image and look at the "M" at the bottom of the middle column. The fewer probes, the more blurring. And the Gaussian probes have more blurring than trilinear probes. This is why we suggest that when efficiency is a concern during animation, non-repeated texture patterns should probably favor parameters that enhance aliasing.
(Not shown in the paper.) High-efficiency Simple Feline with a trilinear probe.
(Not shown in the paper.) High-quality Simple Feline with a Gaussian probe.
(Not shown in the paper.) High-quality Simple Feline with a trilinear probe.
(Not shown in the paper.) Constant-rounding Simple Feline with a Gaussian probe.
(Not shown in the paper.) Constant-rounding Simple Feline with a trilinear probe.
(Figure 22 in the paper.) Mip-mapped EWA paints smooth but sharp text.
(Not shown in the paper) This image color codes the level of detail from which mip-mapped EWA samples texels. The pure blue at the bottom is level 1, the darker blue at the top is level 2.
(Not shown in the paper) Mip-mapped EWA, where the mip-maps are created with a radius 3 Lanczos filter. Note the "blurriness banding" along the line where EWA jumps from using a large ellipse on level 1 to a small ellipse on level 2.
(Not shown in the paper) Mip-mapped EWA, where the mip-maps are created with a box filter. Note how the blurriness banding disappears. We'd be hard pressed to explain why.
(Not shown in the paper) This image color codes the number of probes that Texram uses. The white at the bottom is one probe, the blue at the top is two probes.
(Not shown in the paper) Note the "probe banding" along the line where Texram jumps from using one probe to using two probes.
(Not shown in the paper) This image color codes the number of probes that high-efficiency Simple Feline uses. The white at the bottom is one probe, the blue at the top is two probes.
(Not shown in the paper) Note the "probe banding" along the line where high-efficiency Feline jumps from using one probe to using two probes.
(Not shown in the paper) This image color codes the number of probes that high-quality Simple Feline uses. The white at the bottom is one probe, the pure blue in the middle is two probes, and the darker blue at the top is three probes.
(Not shown in the paper) Note the "probe banding" along the line where even high-quality Feline jumps from using two probes to using three probes. In this image, we caught it at a particularly vulnerable level of detail; the other images above could be made to look as bad or worse, but after awhile you get tired of adjusting parameters.
(Not shown in the paper) This image color codes the number of probes that constant-rounding Simple Feline uses. The pure blue at the bottom is two probes, the darker blue in the middle is three probes, and the darkest blue at the top is four probes.
(Not shown in the paper) Finally, the probe banding disappears! But don't believe constant-rounding Feline isn't vulnerable. You can still see banding in some images if you catch it in the worst-case transition from one probe to two probes, at a level of detail in which the image almost completely blurs to grey. If you really want to eliminate all probe banding, the safest course of action is to use the kludge factors described in Section 3.3, and set them all to 1. This will force Feline to always round up the conservative fProbes to obtain iProbes. Of course, this seriously increases the number of probes.