Skip to content

Locating a title block: bounding boxes, coordinate frames and provenance — a worked example

Companion to the adiro_geometry and adiro_provenance reference pages. It shows, end to end, how a title block's region, its coordinate frame, and an extracted field value fit together — and how the page coordinates are meant to be read. Introduced with #90; the model is shared with the UC-03 geometry work (tracked in #93).

1. The shape, end to end

An extraction pipeline reads an A1 landscape sheet, detects the title-block region, and reads the client field from inside it. In ADIRO that is:

@prefix geom:  <https://w3id.org/adiro/adiro_geometry#> .
@prefix aprov: <https://w3id.org/adiro/adiro_provenance#> .
@prefix md:    <https://w3id.org/adiro/adiro_drawing_metadata#> .
@prefix ex:    <https://example.org/adiro/demo#> .
@prefix xsd:   <http://www.w3.org/2001/XMLSchema#> .

ex:file-A101 a aprov:DrawingFile ;                   # the file the drawing is stored in (a reference, not the bytes)
    aprov:fileName "A101-ground-floor.pdf" ;
    aprov:contentHash "9f2c…" ; aprov:hashAlgorithm "sha256" ;   # identifies the exact bytes
    aprov:pageCount 12 .

ex:page-A101 a aprov:FilePage ;                      # one page of that file: what is true of the page itself
    aprov:belongsToFile ex:file-A101 ;
    aprov:pageNumber 1 ;                             # counted from 1
    aprov:displayWidthPt 2384 ; aprov:displayHeightPt 1684 ;
    aprov:pdfRotationDeg 0 .

ex:sheet-A101 a md:DrawingSheet ;
    md:contains ex:tb-A101 ;
    md:isOnPage ex:page-A101 .                       # the drawing is on this page of this file

ex:frame-A101 a geom:CoordinateFrame ;
    geom:describesPage ex:page-A101 ;                # the frame reads that page…
    geom:coordinateSpace "page_display_pt" ;         # …in one convention: top-left origin, x right, y down
    geom:unit "pt" .

ex:tb-A101 a md:Titleblock ;
    geom:hasBoundingBox ex:bbox-tb-A101 .            # the TITLE BLOCK carries a box…

ex:bbox-tb-A101 a geom:BoundingBox ;
    geom:bboxXYXY "[1900, 1400, 2360, 1660]" ;       # …its region, in the page's frame
    geom:inCoordinateFrame ex:frame-A101 ;
    aprov:hasConfidence 0.97 ;                        # confidence in the BOX (the detector)
    aprov:hasInferenceMeta [ a aprov:InferenceMeta ;
        aprov:inferredWith <https://example.org/models/titleblock-detector@2.1> ;
        aprov:inferredAt "2026-09-20T14:03:00Z"^^xsd:dateTime ] .

ex:fa-client-A101 a aprov:FieldAssertion ;           # a value read from within that title block
    aprov:assertsFieldKind md:ClientField ;
    aprov:hasLiteralValue "Acme Corp" ;
    aprov:capturedCaption "Client" ;
    aprov:assertedBy ex:tb-A101 ;                     # read FROM the title-block region
    aprov:hasConfidence 0.88 ;                        # confidence in the VALUE (a different model)
    aprov:hasInferenceMeta [ a aprov:InferenceMeta ;
        aprov:inferredWith <https://example.org/models/vlm-extractor@5> ] .

Two things are worth reading off this:

  • Only the title block is localised, not each field. hasBoundingBox sits on the Titleblock region; the field values are read from within that one region (assertedBy ex:tb-A101) and are not given their own boxes.
  • Two independent inferences, two confidences. The detector found the box (0.97, via a detector model); a different model read the value (0.88, via a VLM). Each is an InferredEntity, so each carries its own hasInferenceMeta / hasConfidence. A human annotation is an inference step performed by a person, and the same shape with inferredWith omitted and the agent (aprov:inferredBy) a person.

2. Reading the numbers: coordinate frames

bboxXYXY "[1900, 1400, 2360, 1660]" is meaningless on its own — the CoordinateFrame it points at is what makes it interpretable. ADIRO's primary space is page_display_pt: origin top-left, x increasing rightwards, y increasing downwards, in points (1 pt = 1/72 inch). So the box is a rectangle from (1900, 1400) (its top-left corner) to (2360, 1660) (bottom-right) — the bottom-right strip of a 2384 × 1684 pt (A1 landscape) page.

The same rectangle looks different in other spaces, which is exactly why the frame is explicit:

Space Top-left corner of the box How it relates to page_display_pt
page_display_pt (ADIRO primary) (1900, 1400) — origin top-left, y down
Native PDF user space (1900, 284) origin bottom-left: y_native = displayHeight − y_display = 1684 − 1400 (corners also swap in y)
Rendered raster @ 150 DPI (≈3958, ≈2917) px scale by 150/72 ≈ 2.083; DPI changes pixels but not the pt box

Notes:

  • A rendering is a frame of its own. The raster above is a pixel CoordinateFrame with resolutionDpi 150, and a crop of it is a frame that covers a region (coversRegion); see What a model saw (#239).
  • We store page_display_pt as the single primary representation (it matches the extraction pipeline's page geometry). Native-PDF and pixel values are derived views; record them as evidence if a downstream writer needs them, but do not silently interchange the coordinate-space names.
  • Rotation is already baked in. pdfRotationDeg is recorded for interpretation/validation, but page_display_pt coordinates already include the page mapping — do not rotate them again. A page at a different rotation or file version is a different CoordinateFrame.
  • GeoSPARQL, lightly. BoundingBox is a subclass of geo:Geometry so tools recognise it as a spatial shape, but ADIRO does not adopt a GeoSPARQL CRS or spatial-query profile here — WKT serialisation and spatial queries are deferred (#36). Page points are not longitude/latitude.

3. Why this shape

Modelling the box as a first-class, provenance-bearing class (rather than a bare literal on the field) is what lets ADIRO answer "where is the title block, who said so, and how sure" separately from "what does the title block say, and how sure". It also keeps UC-01 and UC-03 on one representation: UC-03 reference symbols will carry BoundingBox / CoordinateFrame the same way, adding only their own geometry roles (#93). The reasoning for reusing GeoSPARQL and for keeping Croissant at the dataset layer is in external ontology imports and the Croissant comparison.