3D Human Behavior Generation through Action and Interaction Synthesis

Abstract

Learning how to share a common space between autonomous systems and humans requires the capability to understand, generate, and forecast human actions and interactions.

This dissertation introduces:

1. Characteristic 3D Poses, semantically meaningful and goal-oriented probabilistic action poses
2. A method to forecast complex sequences of action labels and 3D poses, only requiring 2D observations
3. An approach to generate dynamic human-object interactions from geometry and text.

Defense Presentation

Thesis