Depth of Field Rendering from Sparsely Sampled Pinhole Images

Abstract: Optical lenses cause foreground and background objects to appear blurred, causing an effect called depth of field. Each point in the scene is projected onto the imaging plane as a semitransparent circle of confusion (CoC) with diameter depending on the distance between the point and the lens. In images rendered with a pinhole camera, the entire scene is in focus, but depth of field may be added synthetically for photorealism, aesthetics, or attention guiding purposes. However, most algorithms for depth of field rendering are either computationally expensive or produce noticeable artifacts. This report evaluates two different algorithms for depth of field rendering. Both algorithms are independent of the rendering technique. The first renders only a single pinhole image and uses a light-field based method for image synthesis. The second renders up to 12 pinhole images and uses CoC gathering to create defocus blur. Ideas from both methods are combined in a novel algorithm which uses sparse samples to approximate the light field. Our method produces a closer physical approximation than the other algorithms and avoids common artifacts. However, it may produce ghosting artifacts at low computation times. We evaluate the methods by comparing rendered images to an assumed ground truth generated with the accumulation buffer method. Physical accuracy is measured through structural similarity (SSIM) while artifacts are evaluated through visual inspection. Computation times are measured in the Inviwo software.