Nvidia Corporation

Telefonaktiebolaget Lm Ericsson (publ)

Jacob Munkberg

Jacob Ström

Jon N Hasselgren

Ignacio Llamas

Jim Rasmusson

Christopher Ryan Wyman

Shiqiu Liu

Petrik Clarberg

Tomas Akenine-Möller

Tomas Akenine-Moeller

Tomas Akenine-MÖLLER

The performance of a neural network is improved by applying quantization to data at various points in the network. In an embodiment, a neural network includes two paths. A quantization is applied to each path, such that when an output from each path is combined, further quantization is not required. In an embodiment, the neural network is an autoencoder that includes at least one skip connection. In an embodiment, the system determines a set of quantization parameters based on the characteristics of the data in the primary path and in the skip connection, such that both network paths produce output data in the same fixed point format. As a result, the data from both network paths can be combined without requiring an additional quantization.

Quantizing autoencoders in a neural network

Disclosed approaches may leverage the actual spatial and reflective properties of a virtual environment—such as the size, shape, and orientation of a bidirectional reflectance distribution function (BRDF) lobe of a light path and its position relative to a reflection surface, a virtual screen, and a virtual camera—to produce, for a pixel, an anisotropic kernel filter having dimensions and weights that accurately reflect the spatial characteristics of the virtual environment as well as the reflective properties of the surface. In order to accomplish this, geometry may be computed that corresponds to a projection of a reflection of the BRDF lobe below the surface along a view vector to the pixel. Using this approach, the dimensions of the anisotropic filter kernel may correspond to the BRDF lobe to accurately reflect the spatial characteristics of the virtual environment as well as the reflective properties of the surface.

Reflection denoising in ray-tracing applications

Disclosed approaches may leverage the actual spatial and reflective properties of a virtual environment&#x2014;such as the size, shape, and orientation of a bidirectional reflectance distribution function (BRDF) lobe of a light path and its position relative to a reflection surface, a virtual screen, and a virtual camera&#x2014;to produce, for a pixel, an anisotropic kernel filter having dimensions and weights that accurately reflect the spatial characteristics of the virtual environment as well as the reflective properties of the surface. In order to accomplish this, geometry may be computed that corresponds to a projection of a reflection of the BRDF lobe below the surface along a view vector to the pixel. Using this approach, the dimensions of the anisotropic filter kernel may correspond to the BRDF lobe to accurately reflect the spatial characteristics of the virtual environment as well as the reflective properties of the surface.

Growing community of inventors

Skane, Sweden

Jacob Munkberg