GAD Editorial · Published · · English study · Chinese summary
A door count is only useful when a reviewer can identify which lines were treated as each door. Architectural CAD drawings contain overlapping conventions: an arc can describe a swing, a curve in a fitting or a piece of unrelated annotation. A recognition result needs to preserve enough of the drawing to let someone challenge that interpretation.
FloorPlanCAD is a useful research reference for this problem. Its revised 2021 paper describes an updated dataset of 15,663 CAD drawing blocks across 35 classes, with line-level annotations. The blocks are cropped from source plans rather than representing that many complete buildings. The research task combines identifying individual objects with classifying continuous elements such as walls. Revised paper
The project is historical: its official site reports that it shut down in early 2022. The annotations and website are licensed CC BY-NC 4.0; the authors do not own the drawing copyrights. This article proposes a review method informed by the research; it does not recommend an active supported plug-in. Project status and terms
Begin with a bounded drawing question
For a lawful office evaluation of a separate recognition implementation, choose one drawing revision and one class of object. Ask for the visible door instances in a named zone, for example, rather than a complete building inventory. State whether demolished work, alternative layouts, furniture blocks and external references belong in the count.
Retain the original vector file unchanged. Record the input conversion and any cleanup, including removed text, exploded blocks or clipped geometry. These steps can affect what the system sees. A result on a simplified export should never be presented as a result on the untouched drawing.
Inspect the grouping as well as the label
Ask for a coloured overlay that associates each proposed object with its source lines and a stable identifier. Inspect both the category and the grouping. Two adjacent doors may be merged; one door may be split into several instances. Those errors can cancel out in a total while leaving the inventory unusable.
Use a small manually checked reference set containing ordinary symbols and awkward cases. Include partially clipped objects at the boundary, unusual drawing conventions and crowded intersections. Keep uncertain instances separate from accepted and rejected ones. The acceptance note should explain what evidence resolved each uncertainty.
Carry drawing boundaries into quantities
Avoid adding counts from overlapping crops until duplicates have been reconciled against their positions in the original drawing. Preserve the relationship between each crop and the sheet. A correct count inside a research-sized patch does not establish reliability across an entire drawing set.
Before a quantity enters a cost or procurement workflow, reconcile it with the applicable schedules and a responsible reviewer. Record disagreements as questions about the drawing or classification, rather than silently forcing the totals to agree. This method concerns information extraction, not a statement that the documents are coordinated.
An effective trial ends with a checked instance register, an overlay and a short error log. Report review time and unresolved cases alongside any automated count. The practical value is a traceable inventory whose exceptions are visible, with a clear boundary between research capability and information accepted for a project.