While a two-year-old finds in every stone a world worth examining, in the common datasets that train our most powerful computer vision systems, stones are rarely represented. This constitutes another way of dealing with stones, which could be seen as a form of selected neglect: in the Microsoft COCO dataset (Lin et al., 2015) - one of the most influential training and benchmarking datasets - 80 object categories are defined, and stone is not among them. COCO-stuff (Caesar et al. 2018), an extension of COCO that adds pixel-wise annotations for amorphous, background materials and regions, does include stone as a category, but only as a leaf-level class under the super-category "solid" and again a material within "wall-stone".

​

Computer vision models are typically trained on large image collections that prioritise people, products, vehicles, and recognisable landmarks. Materials such as stones rarely appear as subjects in their own right and are most often treated as undifferentiated background. This omission reflects cultural and institutional assumptions about what is useful, actionable, or valuable for machines to recognise. By constructing a small, situated dataset focused on stones, the project repositions overlooked materials as worthy of sustained attention, drawing an analogy between dominant technological practices and how histories and memories are commonly approached.