ImageSpaceFiltering

Image-Space Filtering with Geometry Preservation

The images below show how image space filtering adapts to and preserves geometry.

This first image shows a model which has a random color-noise shader applied to it. Although the image is not very interesting, it will show the effects of a blur filter very clearly.

This next image shows the same image with a screen-space blur filter applied. In this example, the blur filter is working only within polygon interiors. As you can see, this causes the noise to remain on the edges of all geometry. However, no edge of any geometry is blended with its neighboring surfaces or with the background.

In the next image, the filter was allowed to cross the boundaries of faces as long as they had the same surface normal preserving continuity. This causes much of the remaining noise to disappear. However, any faces along curved surfaces are not sharing colors across their edge boundaries. The effect of this is clearly shown in the image on the right, where each triangle on the curved surface takes on the average color of the noise it contains, but neighboring triangles are not blending together.

No blending across curved face boundaries

In the next image, the filter was allowed to cross the smooth boundaries of faces with a medium tolerance. This causes some of the triangles along the curvature to blend with their neighbors, creating blended rectangular regions.

Some faces have now blended across shared boundaries

In the next image, the filter was allowed to cross the smooth boundaries of faces with a high tolerance. This causes smoothing across all smooth-shaded faces which have shared edges. However, no blending takes place with any surfaces which do not have geometric continuity. This takes the effectiveness of our filter as far as we can go without losing valuable detail. The small percentage of pixels that did not get filtered represent high frequencies inherently, as many of them represent surfaces that are smaller than a pixel.

In the next image, a Gaussian blur filter was used instead of the box blur used in the prior examples. This simply demonstrates that any traditional filter can be applied without compromising the preservation of geometric detail.

In this last image, the same model is shown having been rendered using monti-carlo sampling of an environment light. This image uses the image-space blur described above to eliminate the high frequency noise generated by the monti-carlo sampling. Although I don’t have an image (unfortunately) to show the noise that would have been present otherwise, it is clear in the image below that there is absolutely no visible noise. Most importantly, you can see that the blurring (shown clearly in the prior examples) has had no impact on the clarity of geometric detail. This ability to convolve any particular shader to reduce its high frequency can be applied to any shader. This can be useful for procedural shaders, anisotropic highlight shaders, as well as ray-traced specular or refractive shaders.

SOLAR DIGITAL is a registered trademark and service mark of Solar Digital, Inc.