GAD Editorial · Published · · English study · Chinese summary · Study dated 2026-10-04
A facade inventory sounds like a counting exercise until the team asks what counts as one window. Is a continuous glazed strip one component, several openings, or a collection of frames and panes? AI can accelerate classification only after that question has a useful answer. Otherwise, a precise-looking total simply repeats an uncertain definition.
BuildingNet offers a research starting point. First submitted on 11 October 2021 and accepted at ICCV 2021, it combines labeled exterior building models with a graph neural network that considers relationships between geometric parts. Its dataset includes approximately 2,000 buildings and 292,000 semantic components. Those are research-dataset characteristics, rather than a promise about a practice's facade files. BuildingNet paper and dated record
Translate the label into a project question
Consider a hypothetical practice reviewing a portfolio of existing buildings for a preliminary refurbishment study. The immediate question might be which elevations deserve a closer window survey. That is a narrower and more defensible purpose than producing an order quantity from a classified model.
GAD proposes a label dictionary with three fields: the machine category, the project meaning and the action it supports. A predicted window could trigger a survey task without establishing glazing type, opening mechanism or replacement dimensions. A wall label could group exterior surfaces without identifying their construction or condition.
Record ambiguous examples before selecting software. Include curtain walls, recessed openings, balconies with glazed balustrades and decorative panels that resemble windows. Have two reviewers classify the same small sample. If they disagree, resolve the definition before interpreting disagreement with a model as an algorithmic failure. This exercise exposes weaknesses in the intended inventory itself.
Inspect the partition as well as the name
An accurate category attached to the wrong extent remains problematic. A mesh could divide a single assembly into many pieces or merge several assemblies into one. Therefore inspect both the predicted category and the geometry receiving it. Preserve the original component identifiers so reviewers can locate and correct an error without guessing from a screenshot.
For the pilot, report errors by the architectural consequence. Confusing a canopy with a roof may affect a survey grouping; merging an opening with the surrounding wall may erase an item from the inspection list. Do not allow a large quantity of easy wall surfaces to conceal poor performance on the smaller components that motivated the exercise.
Keep the inventory honest
A useful deliverable has separate fields for predicted category, reviewer decision and unresolved status. Do not silently overwrite the machine result: the corrections are evidence of where the method helps and where it creates work. Keep a small set of difficult examples for regression checks when the model or file-export process changes.
The transfer from a research collection to a particular office's models must be tested. GAD has not run BuildingNet or measured that transfer. The proposed pilot succeeds when a surveyor can understand why an item was flagged and correct it efficiently. It should leave material specification, condition assessment and verified quantities with the appropriate project evidence.