Seminario: Human Body Models

Il seminario sarà tenuto dalla Dott..ssa Silvia Zuffi, ITC-CNR, Milano.

  • Data: 17 dicembre 2015 dalle 11:00 alle 12:00

  • Luogo: Aula 5.5, Scuola di Ingegneria e Architettura, viale Risorgimento 2, Bologna

Silvia Zuffi

Contatto di riferimento:

About the speaker

Silvia Zuffi received a master in electronic engineering from the University of Bologna, and a master and PhD in computer science from Brown University (RI, USA). She has been a postdoc at the Max-Planck institute for Intelligent Systems (Tübingen, Germany) and currently she is a research scientist at ITC-CNR (Milano).

Her research interests include color imaging and reproduction, 2D and 3D shape models, human pose estimation and inference in graphical models.

https://ps.is.tue.mpg.de/people/szuffi

 

Abstract

The human body is one of the most challenging objects to model. It is highly articulated, it deforms with pose changes, it exhibits variability in shape, while clothes and hairstyles create a large variety of appearances. Building accurate body models is of great importance for Computer Vision and Computer Graphics. Our visual environment is largely populated by humans, and we are very efficient in detecting other people, recognizing their identity, estimating pose and motion. We expect automatic systems to do the same, with human performance.

Computer Vision has since its origin focused on such challenging problems, with a long tradition of model-based methods for human detection and pose estimation from images and video. In Computer Graphics the ultimate goal is a faithful reproduction of reality, and creating realistic characters or avatars that populate virtual worlds is one of the major challenges. In Computer Vision we are interested in generative models that allow inferring the variables of interest from data. The number of model variables is usually limited to the quantities of interest, as inverting the generative process is often very hard. In Computer Graphics we are typically not limited from computational constraints. Models are designed to include many variables in order to generate very realistic reproduction of human bodies. Computational constraints play a role only for interactive visualizations.

In this seminar I will introduce models of the human body popular in Computer Vision and Computer Graphics. I will present our recent work to design human body models that are highly realistic, like those used in graphics, but admit a graphical model representation that supports model fitting to data with distributed optimization methods.