Ambient Occlusion Baking via a Feed-Forward Neural Network
Ugo Erra, Nicola Capece, Roberto Agatiello
Dipartimento di Matematica, Informatica ed Economia - Università della Basilicata
Eurographics 2017 - Short Papers Proceedings, Lyon, France, April 24-28, 2017
Abstract
We present a feed-forward neural network approach for ambient occlusion baking in real-time rendering. The idea is based on implementing a multi-layer perceptron which allows a general encoding via regression and an efficient decoding via a simple GPU fragment shader. The non-linear nature of multi-layer perceptions is suitable and effective for capturing non-linearities described by ambient occlusion values. Also, a multi-layer perceptron is random-accessible, have a compact size, and can be evaluated efficiently on the GPU. We illustrate our approach including its quality, size, and runtime speed.
Selected Images
Video
Paper
Links
Neural Network Ambient Occlusion Baking - Source code
Neural Network Ambient Occlusion Baking - Reference implementation
Optix Prime Baking - Source code
Optix Prime Baking - Reference implementation
Acknowledgements
The authors thank NVIDIA’s Academic Research Team for providing the Tesla K40c card under the Hardware Donation Program.
Bibtex
@inproceedings {egsh.20171003,
booktitle = {EG 2017 - Short Papers},
editor = {Adrien Peytavie and Carles Bosch},
title = {{Ambient Occlusion Baking via a Feed-Forward Neural Network}},
author = {Erra, Ugo and Capece, Nicola Felice and Agatiello, Roberto},
year = {2017},
publisher = {The Eurographics Association},
ISSN = {1017-4656},
DOI = {10.2312/egsh.20171003}
}