Probabilistic Models for Multi-Level Scene Understanding

Daphne Koller
Computer Science Department, Stanford University


Wednesday, May 25, 2011
4:30 - 5:30 PM
Y2E2 105

Abstract:

Over recent years, computer vision has made great strides towards annotating parts of an image with symbolic labels, such as object categories or segment types. However, much can be gained from trying to derive a more unified view of the scene, in which we are trying to simultaneously solve multiple vision tasks, in a way that enforces consistency between them. In this talk, I will describe our work in this direction, which uses machine learning to construct richly structured, probabilistic models of multiple scene components. I will demonstrate the value of such modeling for improvements in basic tasks such as image segmentation, object detection, geometric reconstruction, and human pose recognition. The learning of such expressive models poses new challenges, especially when available training data is limited or only weakly labeled. I will describe novel machine learning methods that can train models using weakly labeled data, thereby making use of much larger amounts of available data.

---

Operations Research Colloquia: http://or.stanford.edu/oras_seminars.html