I am building an active appearance model. I have a large training data set. Too big. :/

My computer's RAM cannot handle a model of this size.

If i make a series of smaller PCA models. Is there a way to combine a series of PCA models into one 'averaged' projection space?

Each sample is an image, that is reshaped into a 1D vector.

Vectors are stacked into one large training data matrix, where each row is one sample. We call OpenCV's PCA function and get the 1D eigenvectors and a 2D eigenvalues(1 row per sample) matrix. We are interested in the 1D eigenvector matrix.

Somebody has suggested Structural Equation Modelling. I thought an average of the eigenvectors should work.... provided the components correspond throughout the series of models.

Any clues anyone?