Not so long ago, computers were completely blind. For a machine, any digital image was not a cat, a car, or a human face, but simply an endless table of pixel numbers. Teaching hardware to “see” the real world was considered an almost impossible task. The objects around us constantly change angle, lighting, scale, hide in shadows, or overlap with other things. The ordinary datasets of a few thousand photographs that scientists had were catastrophically insufficient to explain all this visual variety to a machine.
