Provider Comparison¶
Status: PP-DocLayoutV3 preferred for normal proposals; accuracy unvalidated
Authority: Current contract (2026-08-12) and provider implementation defaults
Last reviewed: 2026-08-14
Question¶
What role does each table provider play in Drawing Lab?
Short answer¶
Providers have different jobs. Only PP-DocLayoutV3 is the normal full-page proposer. IBM TableFormer is crop-only structure recovery. The bare Docling Heron checkpoint is not the full Docling pipeline.
Provider roles¶
| Provider | Role | Notes |
|---|---|---|
| PP-DocLayoutV3 | Current preferred full-page proposal provider | Initial recommended threshold 0.50 |
| Geometry detector | Deterministic morphology and grid evidence | Visible fallback; does not auto-start when DocLayout returns nothing |
| PicoDet | Small comparison table detector | Optional comparison; default threshold 0.35 in implementation |
| Docling Heron | Layout proposals for comparison | Bare checkpoint; threshold 0.60 in implementation |
| Microsoft Table Transformer | Independent document-table proposal baseline | Optional comparison; threshold 0.50 in implementation |
| IBM TableFormer | Structure recovery after a table crop exists | Not whole-sheet region detection |
Docling Heron boundary¶
The advanced Heron experiment uses the official Heron layout checkpoint and image processor.
It can expose Table, Form, and Key-Value Region as separate source-labelled proposals.
It does not run Docling page assembly automatically.
It does not run Docling OCR assignment automatically.
It does not run reading-order recovery automatically.
It does not run TableFormer automatically.
Those boundaries keep Heron raw layout evidence comparable and honest.
TableFormer boundary¶
IBM TableFormer operates on a selected table crop. It recovers structure after a region already exists. It does not propose whole-sheet table regions. It does not replace PP-DocLayoutV3 or the shared verifier.
Accuracy boundary¶
Local timing on Apple M1 Max reports PP-DocLayoutV3 inference median approximately 1.27 s.
Complete proposal plus verification median is approximately 1.42-1.48 s per prepared sheet.
Those figures are evaluation evidence.
They are not validated accuracy and not a product SLA.
Open Questions¶
- After representative review, does any comparison provider earn a permanent side-by-side default slot?
Sources¶
- Current contract:
table-dimension-detection-architecture.md(2026-08-12) - Provider defaults in Detection worktree table-region provider configuration