Real-time Character Posing Using Millions of Natural Human Poses

We present intuitive interfaces for interactively posing 3D human characters. The user can create desired fullbody poses by directly dragging points, modifying bone directions, or specifying distances between two points in 2D screen space. Designing such an interface for full-body pose modeling is challenging because many unnatural poses might be consistent with the ambiguous user input. The system automatically learns a pose prior from a huge database which contains about 2.8 million prerecorded poses and uses it to remove the ambiguity. We formulate the problem in a maximum a posteriori (MAP) framework by combining the prior with user-defined constraints. Maximizing the posterior allows us to generate an optimal and natural full-body pose that satisfies the user-defined constraints. Our system runs in real time; it is also simple and easy to use. We evaluate the performance of our approach with cross validation tests and compare with alternative techniques for character posing.