Visual metadata translates an image into searchable statements. That translation is never neutral: it privileges some details, ignores others and depends on the photograph supplied. A collection team can make it more accountable by separating observed form from inferred identity. “Sleeve volume concentrated above the elbow” is a description that another viewer can inspect. “Romantic revival” is an interpretation requiring a wider argument. Both may belong in research, but they should not be stored as if they were the same kind of claim.
Build a description in layers
Start with the broad object category supported by the record. Then describe visible silhouette, length, volume distribution, neckline, sleeve configuration, closures, surface ornament and apparent construction details. Use separate fields where they enable meaningful comparison. A single fluent caption can conceal contradictions and make corrections difficult. “Long” also needs a reference: ankle-length on a mounted figure differs from a measured length in centimetres. Record the observation's basis rather than assuming every photograph provides a reliable scale.
Controlled vocabularies improve consistency, especially across teams. The Getty Art & Architecture Thesaurus provides structured terminology for many cultural heritage concepts. Europeana Fashion offers another point of reference for fashion heritage terminology and aggregation practice. Examine scope notes and language coverage before adopting a term. A vocabulary is not a universal map of dress: retain local names, source-language terms and community-preferred descriptions where a broad label would erase a meaningful distinction.
Describe the image conditions too
A model sees pixels, including the plinth, backdrop, mannequin and conservation supports. Dark fabric against a dark background may hide an edge. A three-quarter view can exaggerate an asymmetric hem. A sleeve can appear full because of its mount rather than its original construction. Record view direction, cropping, visible occlusion and major presentation conditions. When possible, compare front, back and detail views. If only one image exists, the appropriate result may be a limited description, not a confident complete reconstruction.
Do not confuse appearance with substance
A glossy surface is not proof of silk, and a repeated motif is not proof of a particular weaving process. Fibre, weave, construction and finish are related but distinct observations. A photograph may reveal a seam without establishing whether it is original or repaired. Keep machine labels such as “satin-like sheen” provisional and tied to their image. When a conservator or technical report supplies stronger evidence, connect that source to the material field without deleting the earlier visual observation's history.
Make annotation a shared practice
Write a short annotation guide with positive examples, borderline cases and explicit reasons to choose “not visible.” Have more than one person describe a small sample independently, then discuss disagreement before enlarging the dataset. Agreement can indicate a workable instruction, but it cannot prove that the agreed label is historically correct. Reviewers should be able to challenge the vocabulary itself. Repeated disagreement may expose an unsuitable category rather than careless annotation, especially where garments sit between familiar institutional classifications.
Handle cultural labels with care
Appearance alone does not establish an object's cultural identity, maker's identity or community of use. A model trained on unevenly described internet images can reproduce broad, exoticizing or commercially convenient labels. Keep terms from historical catalogues visible as source quotations where necessary, but distinguish them from current descriptive language. Consult relevant expertise when labels concern living traditions or restricted knowledge. Omitting an unsupported attribution is better than turning visual familiarity into a definitive statement about origin.
Export observations without losing meaning
When sharing descriptions, preserve term identifiers, labels, language, source image references and review states. SKOS, a W3C recommendation, provides a useful model for representing concepts and relationships in knowledge organization systems. The DCMI Metadata Terms can support general resource description, but do not replace garment-specific observation fields. A clear local data dictionary is essential even when a recognized standard is used.
Finish by testing whether researchers can recover the observation from the record. Can they identify the exact photograph, understand the descriptor and see why it remains uncertain? If not, richer automated tagging may simply produce more opaque metadata. The purpose is not to generate the largest number of labels. It is to make visible evidence retrievable, comparable and open to correction before any system begins grouping objects into patterns.