Using Deep Recurrent Neural Networks to Build a Motion Search Engine for Animators

Building on top of the numerous recent advances in deep learning (and in machine learning in general), we aim at learning high-level, semantically plausible representations of animation data and human from 3D skeletal data in order to automate or replace different tasks of the animation pipeline which require sometimes rigorous human work.
Specifically, the tasks of interest include automatically classifying and locating human actions inside long, continuous Motion Capture (MOCAP) sequences. This will allow re-usability of the large amount of existing MOCAP data from which it is prohibitively long to retrieve information about the actions it contains.
Another task of interest is the search for movements not based on keywords, but based on similar animations. In this scenario, we aim at being able to find similar existing animations based on a query animation, removing the need for a precise action vocabulary, and again, enabling easy re-use of existing assets.

Faculty Supervisor:

Christopher Pal;Michel Gagnon

Student:

Partner:

Ubisoft Divertissement

Discipline:

Computer science

Sector:

Entertainment and Media; Technology; Other

University:

École Polytechnique de Montréal

Program: