Efficient illumination independent appearance-based face tracking
Image and Vision Computing
| Hi-index | 0.00 |
We present a probabilistic approach to learning objectrepresentations based on the "content and style" bilineargenerative model of Tenenbaum and Freeman. In contrastto their earlier SVD-based approach, our approach modelsimages using particle filters. We maintain separate particlefilters to represent the content and style spaces, allowing usto define arbitrary weighting functions over the particles tohelp estimate the content/style densities. We combine thisapproach with a new EM-based method for learning basisvectors that describe content-style mixing. Using a particle-basedrepresentation permits good reconstruction despitereduced dimensionality, and increases storage capacity andcomputational efficiency. We describe how learning the distributionsusing particle filters allows us to efficiently computea probabilistic "novelty" term. Our example applicationconsiders a dataset of faces under different lightingconditions. The system classifies faces of people it has seenbefore, and can identify previously unseen faces as new content.Using a probabilistic definition of novelty in conjunctionwith learning content-style separability provides a crucialbuilding block for designing real-world, real-time objectrecognition systems.