Combinations of Non-Rigid Deformable Appearance Models
Abstract
A framework for object recognition via combinations of nonrigid deformable appearance models is described. An object category is represented as a combination of deformed prototypical images. An object in an image can be represented in terms of its geometry (shape) and its texture (visual appearance). We employ finite element based methods to represent the shape deformations more reliably and automatically register the object images by warping them onto the underlying finite element mesh for each prototype shape. Vectors of objects from the same class (like faces) can be thought to define an object subspace. Assuming that we have enough prototype images that encompass major variations inside the class, we can span the complete object subspace. Thereafter, by virtue of our subspace assumption, we can express any novel object from the same class as a combination of the prototype vectors. We present experimental results to evaluate this strategy and finally, explore the usefulness of the combination parameters for analysis, recognition and low-dimensional object encoding.
Document Details
- Document Type
- Technical Report
- Publication Date
- Aug 01, 1999
- Accession Number
- ADA366984
Entities
People
- Saratendu Sethi
- Stan Sclaroff
Organizations
- Boston University