Friday, October 31, 2008
3:15 PM
U of R Comp. Studies Bldg. rm 209
Vidit Jain
"Mining Context for Recognizing People in Images"
Identifying a person appearing in an
image is a well-studied, yet far-from-solved problem. One limitation of the existing approaches is that the inference
on a particular image is done independent of the captured scene and any other information available in addition to the
image (e.g., text annotation). In other words, the context associated with the image is often ignored. In this work, we
present methods for extracting two additional cues from an image that facilitate robust recognition of people present
in an image: (1) recognizing the captured event (e.g., a tennis match) to reduce the set of possible identities of the
person in the image, and (2) the co-occurrence statistics of people (e.g., Brad Pitt and Angelina Jolie often appear
together).
We formulate the problem of event classification as probabilistic inference on a novel specification of the
hidden-state random field that selects a subset of the observed features suitable for classification. To address the
second problem of learning co-occurrence statistics, we present a directed graphical model that incorporates an expert
(e.g., a face recognizer for people) in the topic modeling framework to anchor the generated latent topics around
semantics determined by the expert.