Fast and Accurate Depth of Field Rendering
Brian A. Barsky and Todd Jerome Kosloff
Cameras and eyes typically don't image a scene entirely in perfect focus. The region within which objects appear sharp, known as "depth of field," is limited. Computer-generated images without depth of field blurring do not look realistic. In addition, control over what is in focus is useful to control the viewer's attention. For these reasons, we seek to add high quality depth of field blur to computer-generated images. Our goal is to do so without significantly increasing render time.
Our technique is based on a post-process blur filter. Previous depth of field approaches of this type either suffer from artifacts at depth discontinuities or are slow O(N2). We are developing methods that are just as fast as previous fast filters O(N), but do not suffer from artifacts.
A straightforward approach to blurring would involve spreading each input pixel into the surrounding region. However, this is very costly for large blur. We are exploring methods that exploit structure in the point spread function by utilizing a reduced-coordinate representation.
We exploit this compactness to achieve accurate add depth of field blurring in O(N) time, with a constant related only to the complexity (not size) of the point spread function.
Figure 1: Image blurred using one of our methods. In this instance, our method is 38 times faster than previous methods for generating an identical image.