As mentioned in the introduction of this section on data scarcity, the complete lack of training images is not that uncommon a situation, especially in industry. Gathering hundreds of images for each new element to recognize is costly, and sometimes completely impractical (for instance, when the target objects are not produced yet or are only available at some remote location).
However, for industrial applications and others, it is increasingly common to have access to 3D models of the target objects or scenes (such as 3D computer-aided design (CAD) blueprints or 3D scenes captured with depth sensors). Large datasets of 3D models have even multiplied on the web. With the coincidental development of computer graphics, this led more and more experts to use such 3D databases to render synthetic images on which to train their recognition models.
