Abstract
Almost all photorealistic rendering algorithms involve stochastic sampling, which often leads to visually objectionable noise in the image. Avoiding such noise remains one of the most important challenges for photorealistic image synthesis. The nature and quantity of noise depends on the sampling algorithm and number of samples. However, the noise's visibility also depends heavily on the human visual system (contrast sensitivity, masking, etc.). We have developed a novel metric for identifying noise in hyperspectral renderings. Here we used psychophysical noise detection tasks to (1) test how well the metric predicts noise visibility in complex scenes, and (2) determine parameter values that yield acceptable images for a wide range of conditions. The findings not only provide perceptually-based parameter values for detecting noise artifacts with our algorithm, but also yield novel insights into how the visual system distinguishes erroneous artifacts from genuine spatial variations of colour and intensity in the scene. The noise metric uses the information theoretic quantity 'Jensen-Shannon divergence (JSD)', to measure similarities between the spectra of different pixels. During rendering, a threshold parameter determines how much variation in the wavelength distribution is deemed acceptable, at which point no further samples are acquired for that pixel. We created a wide range of scenes containing several textured objects in rooms with spatial variations of lighting and surface material. For each scene, we varied the threshold parameter, yielding a series of images with varying degrees of noise. In a 2AFC task participants identified which of two images matched a noise-free reference rendering. Whole images and patches were tested. Our findings verify that the JSD-based metric correctly predicts human noise detection for rendered scenes and image patches, and identify parameter values that yield acceptable results. We discuss the consequences of this metric for theories of human visual noise detection.
Meeting abstract presented at VSS 2016