Now that we have talked about the basics of the initial algorithm,
let's take a look at some initial results.
The vector graphics image shown on the left was converted to a texture 32x32
pixels in size. The black square indicates a region
approximately 4x4 pixels big
in the texture.
If we filtered the image when down-sampling it by averaging
the region under each texel, we would get a texture as shown in
the first column. This is normally how large
textures or artwork is downsampled to reduce aliasing. Now if we
bilinearly interpolate that
when magnifying it we get something as is shown in left-most image
in the bottom row. You can see that the
texture is extremely blurry.
If we choose, instead, to perform a nearest neighbor sampling when downsampling
the texture, we get something as shown at the top of the middle column. You
can see the colors have not been mixed, but it has a pixelated appearance.
A bilinear filtering of that texture is shown below it. Still, not acceptable.
Finally, if we use the silhouette map algorithm we propose in this
paper, we see the
results in the third column. The top image shows how the silhouette points
define the discontinuities in the 32x32 silhouette map. You can see how
the texture's cells are being deformed to match the edges of the G and of
the star.
The final reconstruction using a silhouette map texture is shown below that.
You can see how this accurately models the magnification of the original
artwork.