Dilapidations Assessment via Computer Vision

At lease end, a dilapidations assessment is a condition comparison against a recorded baseline, not a fresh opinion of the unit in isolation. The surveyor’s job is to separate fair wear and tear from tenant damage, missing items, and unauthorised alterations, then schedule a defensible claim. That work is slow when the move-in record is a folder of photos, a check-in report, and a later set of move-out images that do not line up room by room.

This page describes a computer vision workflow that treats those two image sets as paired evidence. The model does not decide liability, cost, or whether to serve a schedule. It flags visual differences, groups them by space and likely defect type, and hands the surveyor a structured difference list to accept, reject, or recode. The surveyor still writes the schedule, applies the lease and any inventory, and issues the claim.

Related vacate work sits next to this comparison, not inside it. Validity of a break notice is a document question covered in Lease Break Notice Validity Checker. Packing the outgoing file for the landlord or solicitor is covered in Move-Out Document Automation. Finding a replacement occupier is a letting problem, not a dilapidations one: Prospective Tenant Matching Model.

The model needs two image collections that can be aligned to the same property and the same spaces: the move-in baseline (check-in photos, inventory photographs, or a dated condition survey) and the move-out inspection set. Optional context that improves pairing, without becoming a substitute for imagery, includes room labels, floor or unit identifiers, capture timestamps, and a simple space list (kitchen, bathrooms, bedrooms, circulation, exterior if photographed).

Pairing is the first failure point. Photos taken at different distances, with flash versus daylight, or with walls cropped differently, still belong in the same room bucket if metadata or a surveyor-assigned space tag says they do. The model should not invent a baseline from a later visit. If only one side of the pair exists, there is nothing honest to compare.

Empty output is required when move-in imagery is missing, when move-out imagery is missing, or when neither collection can be matched to the same unit. Empty means no difference list, no severity scores, and no implied “all clear.” A blank result is safer than a hallucinated condition. The surveyor then falls back to a full inspection and a manual read of whatever documents remain.

Do not run a silent partial compare on “whatever photos we have.” A kitchen-only move-out set against a full move-in set is still incomplete for a whole-property schedule. Surface coverage gaps as gaps, not as absences of defect.

Once both collections are present and spaces are aligned, the model compares corresponding views: wall planes, floors, ceilings, joinery, sanitaryware, appliances that appear in both sets, and obvious fixtures. Typical flags include new marks, holes, staining, cracked tiles or glass, missing handles or ironmongery, damaged worktops, scuffed skirtings, mould-like discoloration, and items present at check-in that are absent at check-out when both frames show the same location.

Quantification here means measurable or rankable visual change, not a contractor quote. Useful outputs are approximate area of staining or damage on a surface, count of discrete marks, presence or absence of a listed fixture, and a coarse severity band (for example cosmetic, local repair, or likely replacement) that the surveyor can override. Pixel change without a space label is not enough for a schedule line.

The model should keep a conservative stance on wear. Even colour shift and general dulling of paint can be wear rather than breach. Flags that look like uniform ageing, without a localised defect, belong in a “review as wear” bucket rather than an automatic claim item. Tenant damage is easier to support when the baseline shows a clean, intact surface in the same viewpoint.

Lighting, camera angle, and compression create false positives. A flash hotspot can look like a stain. A wide-angle move-out shot can make a scuff look larger. The difference list should therefore carry a confidence note and the paired thumbnails, so the surveyor can dismiss artefacts in seconds instead of re-walking the flat for every alert.

What the model must not do is map a flag straight to a statute, a dilapidations protocol clause, or a pound figure. Costing still needs rates, access, VAT treatment, and whether the landlord will actually reinstate. Those decisions stay with the surveyor and, where instructed, the solicitor.

Human-in-the-loop is the operating rule. The model flags differences. The surveyor still schedules the claim. In practice that means working a queue of paired images: confirm the space, confirm the defect type, drop false positives, merge duplicates from overlapping shots, and add anything the cameras missed (smell of smoke, heating that does not work, keys, meters, loft or cupboard contents not photographed).

Only after that review should items land on a schedule of dilapidations or a check-out report with claimed works. Narrative should stay in the surveyor’s words: lease repairing covenant, inventory exceptions, and any agreement on decoration. The CV layer is an evidence index, not the legal document.

Dispute risk falls when both sides can see the same before-and-after pair for each claimed item. Tenants and managing agents argue less about “that mark was always there” when the check-in frame is attached to the line. The reverse is also true: the surveyor should drop items the model flagged if the baseline already shows the same damage. That cut is as valuable as finding new defects, because over-claiming is what inflates negotiation and expert-to-expert argument.

Time saving is in triage and documentation, not in skipping the inspection. A surveyor still needs to attend when access, safety, or incomplete photography demands it. The gain is fewer hours spent lining up folders, fewer missed comparisons in a large block of similar units, and a cleaner pack if the claim is challenged.

Keep an audit trail: model version, which images were compared, which flags were accepted or rejected, and who signed off the schedule. That trail matters if a later valuer or court asks how an item entered the claim.

Computer vision cannot see what was never photographed. Understairs cupboards, meter cupboards, garden sheds, and the far side of a bath panel are common blind spots. Instruct check-in and check-out photographers (or the surveyor’s own kit) to cover the same spaces, including close-ups of known risk areas: wet rooms, window boards, and kitchen junctions.

It also cannot interpret lease wording. A clause that requires the tenant to decorate in the last year, or that treats certain fixtures as landlord fittings, is outside the image model. Likewise, statutory notices, timing of a terminal schedule, and whether a claim is served as a formal dilapidations package remain professional and legal questions.

Use this comparison when both image sets exist and the question is “what changed.” Use a full traditional survey when imagery is missing (empty output), when the unit was altered so heavily that pairing fails, or when the instruction is valuation or structural opinion rather than lease-end condition. Keep notice checking, file assembly, and re-letting models in their own lanes so a dilapidations file does not get mixed with a marketing match or a break-clause decision.