A Learning-Based Approach For Filtering Monte Carlo Noise

Less noise on a wide range of Monte Carlo effects such as depth-of-field, volumetric rendering, area light sources, and global illumination.  Less costly than traditional ray tracing.  Lower relative mean square errors and sharper images than comparable state-of-the-art algorithms.

Three-dimensional computer graphics Path tracing Video games, architecture designs, computer-generated films, and cinematic special effects

UC Santa Barbara researchers have evaluated the complex relationship between the filter parameters and noisy scene data using a nonlinear regression model, a multilayer perceptron neural network. The network is combined with a filter and trained in an offline process on noisy images of scenes with a variety of Monte Carlo effects. This produces filtered images that are superior to previous approaches in terms of depth-of-field, volumetric rendering, area light sources, and global illumination.

20180114096

Background

Producing photo-realistic images from a model requires taking an intricate multi-dimensional integral of the scene function at every pixel of the image. Generating effects like depth-of-field and motion blur requires computing the integral over different domains, such as lens position and time.  Monte Carlo  rendering systems approximate this integral by sending light rays (e.g. samples) in the multidimensional space to evaluate the scene function, but the inaccuracies of this approximation appear as “noise” in the resulting image.

The most successful approaches in Monte Carlo noise reduction utilize additional scene features such as world position, shading normal, and texture to filter the image. The main challenge with such methods lies in finding the appropriate weights for each feature in the filter. 

Additional Technologies by these Inventors

• Pairwise-Learning Framework for Image Quality Assessment

Tech ID/UC Case

24388/2015-002-0

Related Cases

2015-002-0

USA

Desktop View