Spectral PCA Framework for Pose Estimation

Tom Davis.

An image space representation and indexing scheme intended to serve as the
foundation for a pose parameter recovery algorithm is proposed. The approach
employs a hybrid appearance-based/synthetic rendering technique which uses a
very small set of real object views and a CAD model to generate the ensemble
of all views of an object rotating about one of its stable axes, and it
employs
the principal components of a spatial frequency representation of the
ensemble.
Hence, it is referred to as Spectral PCA.

Spectral PCA is intended primarily for use in industrial parts inspection
and
similar applications, and some fundamental object and observation conditions
typical of that problem domain are assumed. Notably, the camera, observation
table and illumination conditions are assumed fixed, so that the required
algorithm training is adapted to the specific application. A consequence is
severe limits on the training/preparation time requirements if the approach
is to be viable for real applications, and a major focus of the work
reported here is on curtailing the training requirements. The result is a
process
which yields a high fidelity pose recovery algorithm usable over several
inches of translation bounds and requiring about 10 muinutes of training for
a nominal case. Experimental results are provided.