Floor Plan and Architectural Drawing Data for Plan-Reading AI

This page is general information, not legal advice. Confirm requirements with counsel for your jurisdiction and use case.

What public floor plan datasets cover, and where they stop

Public floor plan datasets are strong for residential room segmentation and weak for commercial construction documents. ResPlan offers about 17,000 residential plans as vector graphs with walls, doors, windows, balconies and functional spaces such as kitchens and bathrooms, and its authors argue that CubiCasa5K is too small for current models [1]. MSD (Modified Swiss Dwellings, ECCV 2024) adds more than 5,300 multi-apartment plans covering roughly 18,900 apartments because earlier sets lacked multi-unit layouts [2]. A 2023 set from PUCP contributes 954 high-resolution residential building images with wall and slab polygons [3].

What these sets rarely contain is what a takeoff or plan-review model sees in production. Expect gaps in:

• Discipline coverage: structural, MEP, fire protection and civil sheets, not just A-series floor plans.
• Sheet context: title blocks, sheet indexes, revision clouds, keynotes, schedules and general notes.
• Drawing noise: hatching, overlapping dimension strings, xrefs, scanned redlines and mixed scales on one sheet.
• Commercial typologies: retail, healthcare, warehouse and tenant-improvement plans with grid lines and demolition phasing.

Before training, check each public set's license on its own repository or card [8]; academic availability does not settle commercial use. Our guide to public document datasets and commercial use covers how to read those terms.

Which labels takeoff, segmentation and plan-review models need

The label schema should follow the downstream task, because room polygons alone cannot drive quantity takeoff. Room segmentation needs closed room polygons with type labels. Takeoff needs wall centerlines with thickness and type, opening instances tied to door and window schedule tags, and a verified scale so pixel lengths convert to feet and inches. Plan review needs sheet metadata, code-relevant annotations and links between reviewer comments and drawing regions.

Illustrative example: invented to show structure; it does not describe an available dataset.

Store geometry in a documented format (COCO-style JSON for raster masks and boxes, or a vector graph keyed to source DWG, DXF or PDF coordinates) and keep one coordinate system per sheet. If you also extract text, align it to the conventions on our document annotation schema page and our OCR ground truth guidance so title-block and keynote text is consistent with the rest of your document stack.

How sheet index and title block extraction differs from form extraction

Title block extraction is a key-value problem with drawing-specific traps. Field positions vary by firm template, revision tables grow over a project, and the sheet number on the title block must match the sheet index on the cover sheet. Labels should capture sheet_number, sheet_title, discipline prefix (A, S, M, E, P, FP), revision number and date, issue purpose (for permit, for construction, addendum) and project phase.

For mechanical and manufacturing drawings, where GD&T and part-level dimensions dominate, see engineering drawing annotation data; generic form approaches are covered in key-value extraction labels. Architectural plan sets add one more requirement: keep every revision of a sheet as a separate, dated record, because models that read only the final set never learn what changed between bulletins.

Where plan-review comment data comes from

Plan-review training data pairs drawings with reviewer comments, and public examples exist but are rarely linked to drawing regions. Seattle publishes a Plan Comments dataset of reviewer remarks on building permit plan sets, typically many per plan set, under an "other license specified" designation that needs reading before commercial use [4]. Comments in such feeds usually reference a sheet or a code topic in free text rather than a polygon on the drawing.

For a supervised plan-review agent, you need comment-to-region links, the code section cited, the cycle number and the resolution (revised, clarified, deferred). That linkage normally exists only inside an architecture firm's or reviewer's own markup tools and correspondence. Project-level records such as RFIs, submittals and logs belong to construction project datasets; this page covers the drawing layer.

Who holds rights in architectural plans, and what to check

Rights in plan sets are split among the architect, consultants and the building owner, so check every project rather than assuming the holder of the files can license them. Owner-architect agreements commonly leave copyright in drawings with the design firm [6] while granting the owner use rights for the project, and consultant sheets (structural, MEP) may belong to the consultants. Apart from transfers by operation of law, a transfer of copyright ownership is invalid without a signed writing [5], so if a seller says rights moved to it, ask for the signed instrument; a verbal "we own our drawings" is not evidence.

Practical checks for buyers:

1. Ask which entity authored each discipline and whether consultant agreements allow reuse beyond the project.
2. Confirm the owner-architect agreement does not restrict reuse of instruments of service or require client consent.
3. Exclude or treat separately sheets for sensitive facilities (security plans, critical infrastructure, data centers).
4. Mask street addresses, owner and tenant names, permit numbers and stamped seals with license numbers in title blocks.
5. Record the masking method; NIST notes that traditional de-identification has limits [6], and a floor plan with a distinctive footprint can still point to a building.

For contractor-drawn or outsourced annotation, verify IP assignment as described in contractor-created data and IP assignment.

A request template for licensed plan set data

A precise request describes drawings, labels and rights, not the firms that might hold them.

Illustrative example: invented to show structure; it does not describe an available dataset.

request: architectural_plan_sets
typologies: [tenant_improvement, healthcare_outpatient, light_industrial]
disciplines: [A, S, M, E, P]
source_formats: [vector_pdf, dwg]          # scanned TIFF acceptable for redline subset
phases: [permit, construction, addenda]
revisions: all_issued_with_dates
labels_required:
  - rooms: polygon + room_type
  - walls: polyline + type_tag + phase
  - openings: bbox + schedule_ref
  - title_block: sheet_number, discipline, revision, date
  - scale: px_per_ft verified per viewport
optional:
  - plan_review_comments linked to sheet + region
deidentification: mask addresses, owner/tenant names, seals, permit numbers
exclusions: security and critical-infrastructure sheets
acceptance: 5% sample audit, wall-length error and room IoU thresholds agreed
documentation: dataset card + Croissant-RAI fields

Documentation should travel with the data: a dataset card with license and label definitions [8], and machine-readable responsible-AI fields covering collection, labeling and limitations [7]. Use our annotation quality audit method to set acceptance thresholds before delivery.

How SourceX approaches plan set sourcing

SourceX sources operational datasets from US companies on request, including documents and engineering records, and manages the licensing process. Datasets are not held in stock, so a request for plan sets is a search, not a catalog order, and it does not guarantee a match. Every release is approved by the supplying company, each dataset is rights-reviewed for ownership and consents, and personal details are removed or replaced before delivery with the method recorded. Buyers can describe their target drawings at SourceX for AI data buyers; the architecture firms page and CAD and PCB design datasets cover adjacent sources. More document tasks are listed in the Document AI data hub.

Source floor plan and drawing data for your model

SourceX looks for US businesses that hold the plan sets you describe and runs Find, Assess, Agree, Transact and Manage, with nothing contracted until a supplier agrees. Each dataset is delivered under a license defining records, uses, term and delivery. Describe your drawing and label requirements at SourceX for AI data buyers.