Projects

Data-driven 3D Modeling

Making three-dimensional content creation easier is one of the main challenges in computer graphics. While expressive modeling tools have been developed for professional digital artists, people who have not undergone extensive training remain unable to create compelling 3D objects and environments. The goal of this project is to enable the general public to create detailed three-dimensional content. Our approach is to develop intelligent 3D modeling tools that maintain an internal representation of the space of shapes that can be produced. On the one hand, this representation should be rich enough to yield a comprehensive variety of desirable shapes. On the other hand, a casual user should be able to navigate the space of shapes and arrive at one that suits his or her needs. We have explored a number of representations and associated modeling interfaces. Continue reading →

Procedural Modeling

Procedural modeling is an approach to three-dimensional content creation that describes shapes through generative representations that specify not only the final product, but the procedure that gives rise to it. Procedural representations are particularly effective for describing complex shapes that exhibit detailed structure on multiple scales. There is a long history of procedural specification of natural phenomena, such as plants, landscapes, and clouds, but recent work has begun to seek out procedural representations for man-made objects. Unfortunately, procedural representations are often difficult to author and their output can be difficult to customize. We are developing techniques designed to overcome these challenges. Continue reading →

Dense Probabilistic Models for Computer Vision

This project explores efficient algorithms for densely connected probabilistic models, and applications of such models in computer vision and related areas. We have introduced a highly efficient algorithm for inference in densely connected random fields. The algorithm enables rapid inference in conditional random fields that connect every pixel in an image. We have shown that such densely connected models lead to substantially improved image understanding. Continue reading →

Learning Behaviors from Demonstration

This project explores novel approaches to authoring the behavior of animated characters. We would like to enable people without programming expertise to populate interactive three-dimensional environments with characters that appear intelligent and behave according to the author’s intent. Our approach is to allow non-specialists to specify the behavior of animated characters by acting out examples of the behavior. To this end, we have developed new algorithms for inverse reinforcement learning, which is a core technical component of example-based behavior authoring. Continue reading →

Gesture Animation

Body language forms a crucial component of face-to-face interpersonal communication. However, it has been conspicuously missing from interactions in networked virtual environments. While voice communication in virtual worlds has become widespread, gestural communication is usually performed with manual keyboard control or not at all. The goal of this project is to endow avatars in virtual worlds with expressive body language that corresponds to the speech of the operators. Continue reading →