Abstract: We will first
present a new class of model-based learning methods which include hypergraph
and structured sparse learning for vision understanding. In our hypergraph framework,
a hyperedge is defined by a set of vertices with similar attributes. The
complex relationship between the mages can be easily represented by different
hyperedges according to different visual cues. We applied unsupervised and
semi-supervided hypergraph learning to video object segmentation and image
retrieval. Extensive experiments demonstrate its advantages over traditional
graph models. Our structured sparsity framework is a natural extension of the
standard sparsity concept in statistical learning and compressive sensing. By
allowing arbitrary structures on the feature set, this concept generalizes the
group sparsity idea. A general theory is developed for learning with structured
sparsity, based on the notion of coding complexity associated with the
structure. We will show applications in sparse learning, compressive sensing,
computer vision and medical imaging. Time permitting we will show applications
of stochastic deformable models to facial tracking, body tracking, ASL
recognition and cardiac motion analysis/simulation.
- This event has passed.