How would you go about aggregating an image library to properly gauge the classification results? Would you just use ImageNet content or try using something ent

How would you go about aggregating an image library to properly gauge the classification results? Would you just use ImageNet content or try using something entirely new?