PDF version of this entire document
To avoid over/under-fitting, size of database matters a lot and our performance depends on the quality of the data used for training. We could try just running it on the NIST dataset, but it would not reproduce what the IJCV paper states - a paper wherein they basically worked on 3,000+ 3-D images taken from 3 individuals at their lab. They trained their statistical models based on this vast dataset before applying it to an unseen test set.
Finding a conceivable way to access many date files or pass them to/mount them on the computational server (as done with NIST data). The program is ready to run experiments, it just needs 3-D data from different individuals.
