Skip to content

Table Detection Overview

Status: PP-DocLayoutV3 selected for normal proposals; representative accuracy unvalidated

Authority: Current contract (2026-08-12)

Last reviewed: 2026-08-14

Question

How does Drawing Lab find and review tables without claiming finished understanding?

Short answer

A region provider proposes a possible table. A shared verifier checks finite rules and compatible evidence. A human accepts, adjusts, rejects, or adds a missing region. Later stages recover structure, assign OCR to cells, and understand fields.

flowchart TD
    A[Whole-sheet raster] --> B[Region provider proposals]
    B --> C[Shared verifier]
    C --> D[Append-only human review]
    D --> E[Rule and structure recovery]
    E --> F[Cached OCR assigned to cells]
    F --> G[Understand type headers values fields]

What PP-DocLayoutV3 receives and returns

PP-DocLayoutV3 receives the canonical whole-page Detection raster. It returns source-labelled boxes whose label contains table. Each box carries a provider score at the initial recommended threshold 0.50. The proposal keeps page coordinates and starts in needs_review. The provider does not assign table type, headers, or cell text.

Proposal confidence versus verifier evidence

Provider confidence stays visible as a diagnostic. Provider confidence cannot amplify verification. The shared verifier measures rules, intersections, OCR support, and negative evidence. Auto decisions use measured verifier evidence, not the model score alone.

Shared verifier in plain language

The verifier is provider-neutral. Every comparison provider sends boxes through the same checks.

Finite rule segments are short horizontal and vertical separators found in ink. Measured intersections are real crossings of those separators. Provisional cells are early cell candidates before full confirmation. Cached OCR compatibility checks whether existing readings fall inside the proposal. Negative evidence looks for drawing-like ink, arrowheads, arcs, or an oversized box. Weak verification can leave a proposal uncertain. An uncertain or weak proposal remains reviewable. Human review still accepts, adjusts, rejects, or adds missing truth.

Current route

  1. PP-DocLayoutV3 proposes regions at threshold 0.50.
  2. Geometry remains a visible fallback.
  3. PicoDet, Docling Heron, and Microsoft Table Transformer remain optional comparison sources.
  4. IBM TableFormer recovers structure from a selected table crop only.
  5. Providers are never automatically fused.
  6. Accuracy remains unknown until representative fully reviewed pages are measured.

Human review and later stages

Accepted or corrected regions become append-only truth. Missing tables can be added by hand. After region review, a structure pass finds finite separators and cells. Cached OCR is then assigned to the selected table. Understand later assigns type, headers, values, and title-block fields.

Claim boundary

Do not claim validated table accuracy. Current evidence supports a preferred provider choice. Current evidence does not support a production accuracy score. Review-rail counts are practical proxies, not formal precision or recall.

Open Questions

  • Which representative pages open the first formal region precision and recall report?

Sources

  • Current contract: table-dimension-detection-architecture.md (2026-08-12)
  • Local timing evaluation on Apple M1 Max in the same contract