Choose the unit of analysis
A whole building, a facade, a roof and an individual doorway are not interchangeable training examples. A doorway may have been salvaged; a facade may conceal an older structure; a roof may have been rebuilt. Name the unit in every record. If several units belong to one building, give them linked identifiers rather than silently treating each photograph as an independent building.
This matters when evaluating a system. Photographs of the same structure can look different while still sharing distinctive features. If one photograph enters training and another enters testing, the reported performance can reward recognition of a familiar building rather than generalization. Split related records by building and consider site-level grouping when several structures share a construction campaign or photographic setting.
Write categories that can be challenged
A category definition should state inclusion rules, exclusion rules and examples of ambiguity. “Classical” is too broad unless the project explains whether it covers an order, a composition, a historical movement or a later reference. Do not assume terminology travels unchanged between languages or regions. Preserve original terms beside normalized labels, and maintain a glossary that records the project’s choices.
Use a hierarchy when evidence supports different levels of specificity. A model may recognize an arched opening reliably while failing to distinguish its period. Permit a broader label instead of forcing a narrow one. Multi-label records can represent mixed features, provided each label has its own supporting evidence. “Unknown,” “not visible” and “outside scope” should remain distinct outcomes rather than becoming a single miscellaneous class.
Use reference buildings carefully
Cologne Cathedral, the Bauhaus building in Dessau, Villa Savoye in Poissy and the Alhambra in Granada offer recognizable points of comparison. They do not constitute a representative dataset of European architecture. Famous monuments are photographed more often, documented more thoroughly and presented from more carefully selected viewpoints than ordinary streets or altered rural buildings. A model trained mainly on such examples may learn a catalogue of icons.
Include less celebrated buildings, later additions, vernacular forms and difficult cases. Record why a reference label was accepted: a scholarly publication, a survey, a conservation report or an official inventory may support different claims. Agreement among copied web descriptions is not independent corroboration. If the reference label is disputed, preserve that dispute rather than converting it into an unquestioned training target.
Separate research purpose from designation
UNESCO’s World Heritage framework concerns a particular international designation; it is not a universal classifier of architectural style. In France, the Ministry of Culture’s POP platform provides access to heritage resources, while protection and documentation still require careful reading of the relevant records. In Germany, monument protection is governed at state level. A visual resemblance cannot substitute for checking the responsible authority and the exact property or component covered.
Before collecting a large dataset, write a short scope statement that another researcher can test. Specify the geography, building types, image conditions, category vocabulary and acceptable uncertainty. Describe who may review labels and when the system must abstain. This small document becomes the benchmark for later decisions: an impressive prediction outside the defined scope is still outside scope.