Pixar

Disney Enterprises, Inc.

Aurora Operations, Inc.

Vis Machina Inc.

Disnev Enterprises, Inc.

Alex Harvill

Mark Joseph Meyer

Jan Novak

Brian McWilliams

Thijs Vogels

Fabrice Rousselle

Michael Fu

Allen Hemberger

A modular architecture is provided for denoising Monte Carlo renderings using neural networks. The temporal approach extracts and combines feature representations from neighboring frames rather than building a temporal context using recurrent connections. A multiscale architecture includes separate single-frame or temporal denoising modules for individual scales, and one or more scale compositor neural networks configured to adaptively blend individual scales. An error-predicting module is configured to produce adaptive sampling maps for a renderer to achieve more uniform residual noise distribution. An asymmetric loss function may be used for training the neural networks, which can provide control over the variance-bias trade-off during denoising.

Temporal techniques of denoising Monte Carlo renderings using neural networks

Virtual creation and visualization of a physically based virtual material swatch includes identifying a capture device model to support for use in the generation of the virtual material swatch, determining camera response of the capture device model, determining lookup tables of linearly transformed cosines for a plurality of parameters of a shading model, identifying a sample physical material, and determining material description for the sample physical material. A system and method may also include scanning a user physical material, selecting one or more of material preset, geometry preset, environment preset, and camera preset, and creating a virtual material swatch for the user physical material based on the selections.

Virtual creation and visualization of physically based virtual materials

A computer-implemented method of performing machine vision prediction of digital images using synthetically generated training assets comprises digitally capturing a plurality of assets; configuring each of the assets in the plurality of assets with a plurality of asset attributes; under computer program control, selecting a plurality of different combinations of parameters from among the plurality of asset attributes, and creating a plurality of sets of different synthetic dataset parameters; using computer graphics software, and example parameter values from among the synthetic dataset parameters, creating a synthetic dataset by compiling from a plurality of example images and metadata; configuring a plurality of machine learning trials and executing the trials to train a machine vision model, resulting in creating and storing a trained machine vision model; executing a validation of the trained machine vision model; and inferring a prediction using the trained machine vision model. Trained models are scored against success criteria and re-trained using pseudo-random sampling of different parameters clustered around failure points. As a result, machine vision models may be trained with high accuracy using large datasets of synthesized digital images that are richly parameterized, rather than human captured digital images.

System and method for capture and adaptive data generation for training for machine vision

Adaptive sampling in Monte Carlo renderings using error-predicting neural networks

Denoising Monte Carlo renderings using neural networks with asymmetric loss

Multi-scale architecture of denoising monte carlo renderings using neural networks

Growing community of inventors

Berkeley, CA, United States of America

Alex Harvill