Physics-based Animation and Machine Learning at the Computer Graphics Lab - ETH Zurich/
Azevedo, Vinicius
Physics-based Animation and Machine Learning at the Computer Graphics Lab - ETH Zurich/ Seminário de Computação Gráfica: Vinicius Azevedo. - Rio de Janeiro: IMPA, 2019. - video online
Abstract: The creation of physically-based simulations for high end animations in feature films or visual effects in computer games is an inherently iterative process: the constant interplay between artists and digital content requires tools to change the scene setup iteratively, and efficient computations are needed to avoid long waiting times. In this talk I will present some algorithms and techniques that allow a better integration of physically based simulations into the workflow of artists, particularly focusing on computational efficiency and artistic controllability. The foundation of our work is based on data-driven concepts: we leverage machine learning to improve existing simulations as well as to enable novel applications. First, I will present an efficient method for synthesizing fluid simulations from a set of reduced parameters with deep learning techniques. In the second part of the talk, I will present a method to transfer artistic styles to fluid simulations in a machine learning manner. Lastly, I will conclude with a few open problems in the simulation of fluids and further interesting directions for exploration in integrating Machine Learning with Physically-based Animation .
Matematica.
Physics-based Animation and Machine Learning at the Computer Graphics Lab - ETH Zurich/ Seminário de Computação Gráfica: Vinicius Azevedo. - Rio de Janeiro: IMPA, 2019. - video online
Abstract: The creation of physically-based simulations for high end animations in feature films or visual effects in computer games is an inherently iterative process: the constant interplay between artists and digital content requires tools to change the scene setup iteratively, and efficient computations are needed to avoid long waiting times. In this talk I will present some algorithms and techniques that allow a better integration of physically based simulations into the workflow of artists, particularly focusing on computational efficiency and artistic controllability. The foundation of our work is based on data-driven concepts: we leverage machine learning to improve existing simulations as well as to enable novel applications. First, I will present an efficient method for synthesizing fluid simulations from a set of reduced parameters with deep learning techniques. In the second part of the talk, I will present a method to transfer artistic styles to fluid simulations in a machine learning manner. Lastly, I will conclude with a few open problems in the simulation of fluids and further interesting directions for exploration in integrating Machine Learning with Physically-based Animation .
Matematica.