Modeling Appearance Patterns in Image Sets
par/by Matt Toews
Centre For Intelligent Machines
McGill University
The recent proliferation of imaging devices such as digital cameras or medical scanners has produced masses of image data. Organizing this image data in an intuitive manner requires a means of describing or modeling the appearance patterns they contain, such as faces or places. In order to be practically useful, such appearance models must lend themselves to efficient learning from sets of natural, noisy images with minimal supervision, and to robust inference in order to detect and localize pattern instances in new images.
In this talk, I will describe a new probabilistic model of pattern appearance well suited for learning and inference in difficult imaging conditions. The model describes the appearance of image patterns in terms an OCI (object class invariant), a geometrical structure that is 1) uniquely defined in all instances of an image pattern and 2) invariant to image deformation arising from the imaging process. By modeling an OCI relative to robust invariant image features, image patterns can be efficiently learned from large sets of natural imagery containing clutter, occlusion, and significant pattern appearance variation due to illumination change, geometrical deformation and multi-modal appearance variability (i.e. a face with/without sunglasses). Once learned, the OCI model can be used to efficiently infer pattern instances in new images, in similar difficult imaging conditions. I will give an overview of the OCI modeling approach, including model learning and inference, and I will show how the approach can be used to model faces in digital camera imagery and brains in magnetic resonance imagery.
* This is joint work with Tal Arbel from the Centre for Intelligent Machines and Louis Collins from the Montreal Neurological Institute.