A model trained to detect cracks in concrete fails on brick. One trained on highway pavement fails on bridge decks. The failure isn't computational — the models learn well from their training data. The failure is ecological: cracks look different in different materials, under different lighting, at different scales, in different climates. No single dataset captures this variation.
StructDamage (arXiv:2603.10484) aggregates 78,093 images from 32 publicly available sources, reannotated under a unified protocol across nine surface types: walls, tile, stone, road, pavement, deck, concrete, brick, and asphalt. The images span geographic contexts from arid to tropical, historic buildings to modern infrastructure.
Note (2026-11-10): I checked the abstract and searched the full text of arXiv 2603.10484. The dataset has about 78,093 images from 32 publicly available datasets, reannotated and organized in a folder-level classification hierarchy; the abstract says nine surface types but lists eight (walls, tile, stone, road, pavement, deck, concrete, brick), so the essay's ninth, asphalt, is its own addition, and it reports baseline results from fifteen architectures (best DenseNet201 at 98.62 percent accuracy). I did not find the essay's range from arid to tropical climates or its account of the original datasets using bounding boxes, pixel masks and binary labels in the text I searched.
The unification is the contribution. Each original dataset used its own annotation scheme, resolution, labeling granularity, and organizational format. A crack in one dataset might be a single bounding box; in another, a pixel-level segmentation mask; in a third, a binary classification label. StructDamage standardizes them into a folder-level classification hierarchy compatible with modern architectures.
The structural insight: infrastructure monitoring fails not because detection algorithms are weak but because they can't generalize. Generalization requires diversity in the training data, and diversity in infrastructure is enormous — nine material types, hundreds of failure modes, millions of environmental conditions. The problem isn't building a better crack detector. It's building a crack detector that knows what it hasn't seen. The atlas doesn't contain every crack. It contains enough variety that the model can learn what crack-ness looks like across contexts, rather than memorizing what one crack looked like in one photo.