ORBDetector, matchBruteForce, sort and findHomography

My goal is to serialize the needed descriptors and keypoints (found with ORBDetector.detect()) for a source that will be used to align many incoming images. I'll use that source data with incoming images to align them. My thought is that rather than needing the original image for every pass, I hope that serializing it and keeping in memory/cache will be faster. Bad assumption?

Anyhow, I can't seem to reconstitute the descriptor. .sizes looks to be a simple array, but it doesn't come across with new cv.Mat( serializedSource.descriptors ) when serializedSource.descriptors.sizes is an array with just 2 values ( keyPoints * bits from the looks of it ).

I know a Mat has image data behind it typically, but then why the "descriptor" if it's "just a mat", so I'm assuming it's more like meta needed with the keyPoints and I should be able to serialized, saving space this way.

Am I all wrong in my idea here? Should I just load up that source image and detectOrb and then persist that data every time?

Thanks in advance and doubly-so since this isn't a very clear "issue" or problem :)