I have a training set of images, for each of which I\'ve detected and computed their feature vectors (using ORB feature descriptors and extractors. The questions is<
I would suggest storing the feature vectors as images to have a simple and compact representation. You could even use non-destructive compression such as png to minimize file size.