Abstract:

High-level object class recognition involves both the task of 
classification and that of semantic localization (identifying an 
object's outline). While many recent works focus on classification, 
the task of localization has received much less attention. In this 
talk I will present a unified approach for addressing these two tasks 
in concert, and focus on generic machine learning tasks that arise in 
this context.

I will first pose object outlining as probabilistic inference based 
on an explicit shape model. To cope with the computational 
complexity of this problem, I will develop a template algorithm that 
significantly improves the performance of an important and popular 
family of approximate inference methods. I will show how our 
inference-based approach achieves classification rates that are 
competitive with state-of-the-art methods while at the same time 
provides us with accurate outlines that are semantically meaningful.

I will then consider the challenges of learning with weaker 
supervision and with few samples. I will present an approach that 
builds on the tools used for object outlining to learn semantic shape 
models from simple outlines, and show that the models learned are 
competitive with those learned with full supervision. Finally, I will 
present a general purpose hierarchical framework for transfer 
learning between joint distribution representations of related 
classes. I will show how this approach allows us to learn better 
class-specific shape models from few samples by learning jointly 
from several related quadruped classes.