Keep observation separate from diagnosis
Computer-vision systems are often trained to output convenient names: crack, scratch, dent or corrosion. Those names can hide a difficult labelling problem. A visible line is a feature in an image; a crack is a physical discontinuity in material. The first does not automatically establish the second. A useful inspection record therefore keeps two fields: an observed description and, if available, a conclusion made by an appropriately qualified reviewer.
For example, “pale line crossing the painted surface near the frame junction” preserves what the image supports. “Frame failure” adds a structural conclusion. The distinction is not timid wording. It prevents a later user from treating an unverified training label as an established fact.
Surface wear depends on where it occurs
Abrasion on a replaceable grip and abrasion on a load-bearing strap are different problems even when the photographs look similar. Location, material thickness, construction and intended function all matter. A model that recognises texture may not know which layer carries load or whether a visible coating is protective. It may also confuse ordinary manufacturing texture with wear if that texture was absent from its training set.
Record whether the feature is on a contact patch, an edge, a joint or an attachment point. Where wear indicators exist, consult the exact manufacturer instructions rather than estimating a limit from appearance. Schwalbe and Continental publish tyre information, but tyre construction and replacement guidance vary by product. A general image label cannot stand in for that product-specific guidance.
Lines need more than a confident name
Paint boundaries, mould seams, decorative graphics and reflections can all create narrow high-contrast lines. Change the lighting and view direction before interpreting the pattern. A line that remains visible in multiple images is better documented, but persistence alone does not prove a crack. Conversely, a line disappearing under glare is not evidence that the underlying concern has gone away.
Do not scratch away paint or apply improvised penetrants to settle the question. These actions may alter the evidence or damage the item. If a structural crack is suspected in a frame, racket or protective component, stop using the item pending appropriate guidance. The practical decision to pause use can be justified by uncertainty without pretending the photograph has confirmed a failure mechanism.
Shape changes require reference geometry
A wheel can look oval in an angled image. A curved tube can look dented where a reflection changes. To document apparent deformation, keep a view that shows the whole component and another that shows the local feature. Compare only like-for-like views where possible. Do not straighten, flex or load the component to improve the photograph.
Even a real dent has no universal photographic severity threshold. Material, wall thickness, location and service conditions affect its significance. An algorithm trained on metal panels cannot be assumed to assess a composite racket or a foam helmet. Similar image features do not imply similar engineering consequences.
Material changes can conceal deeper uncertainty
Fraying, exposed fibres, flaking, discolouration and corrosion products deserve specific descriptions. Name what you see rather than inferring depth. A fibre visible at a composite surface does not reveal the state of every layer beneath it. Rust-coloured staining does not measure remaining section thickness. A lifted edge may be a label, a coating or a bonded layer; the photograph alone may not resolve which.
For climbing and protective equipment, use the manufacturer’s inspection resources. Petzl, Black Diamond and C.A.M.P. provide product and safety information. Their names here identify source organisations, not a shared inspection rule. Product category, age, use history and retirement criteria must be checked individually.
Allow uncertainty in the label system
A practical dataset should permit “uncertain,” “not assessable” and multiple observations on one area. Forcing every crop into “good” or “damaged” rewards false certainty. Keep benign lookalikes in the review process and record why an expert changed a label. When no suitable reference assessment exists, do not silently promote a visual guess into ground truth.